Daily Papers

Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) have demonstrated impressive language/vision reasoning abilities, igniting the recent trend of building agents for targeted applications such as shopping assistants or AI software engineers. Recently, many data science benchmarks have been proposed to investigate their performance in the data science domain. However, existing data science benchmarks still fall short when compared to real-world data science applications due to their simplified settings. To bridge this gap, we introduce DSBench, a comprehensive benchmark designed to evaluate data science agents with realistic tasks. This benchmark includes 466 data analysis tasks and 74 data modeling tasks, sourced from Eloquence and Kaggle competitions. DSBench offers a realistic setting by encompassing long contexts, multimodal task backgrounds, reasoning with large data files and multi-table structures, and performing end-to-end data modeling tasks. Our evaluation of state-of-the-art LLMs, LVLMs, and agents shows that they struggle with most tasks, with the best agent solving only 34.12% of data analysis tasks and achieving a 34.74% Relative Performance Gap (RPG). These findings underscore the need for further advancements in developing more practical, intelligent, and autonomous data science agents.

Existing online benchmarks for mobile GUI agents remain largely app-centric and task-homogeneous, failing to reflect the diversity and instability of real-world mobile usage. To this end, we introduce VenusBench-Mobile, a challenging online benchmark for evaluating general-purpose mobile GUI agents under realistic, user-centric conditions. VenusBench-Mobile builds two core evaluation pillars: defining what to evaluate via user-intent-driven task design that reflects real mobile usage, and how to evaluate through a capability-oriented annotation scheme for fine-grained agent behavior analysis. Extensive evaluation of state-of-the-art mobile GUI agents reveals large performance gaps relative to prior benchmarks, indicating that VenusBench-Mobile poses substantially more challenging and realistic tasks and that current agents remain far from reliable real-world deployment. Diagnostic analysis further shows that failures are dominated by deficiencies in perception and memory, which are largely obscured by coarse-grained evaluations. Moreover, even the strongest agents exhibit near-zero success under environment variations, highlighting their brittleness in realistic settings. Based on these insights, we believe VenusBench-Mobile provides an important stepping stone toward robust real-world deployment of mobile GUI agents. Code and data are available at https://github.com/inclusionAI/UI-Venus/tree/VenusBench-Mobile.

We present ExAct, a new video-language benchmark for expert-level understanding of skilled physical human activities. Our new benchmark contains 3521 expert-curated video question-answer pairs spanning 11 physical activities in 6 domains: Sports, Bike Repair, Cooking, Health, Music, and Dance. ExAct requires the correct answer to be selected from five carefully designed candidate options, thus necessitating a nuanced, fine-grained, expert-level understanding of physical human skills. Evaluating the recent state-of-the-art VLMs on ExAct reveals a substantial performance gap relative to human expert performance. Specifically, the best-performing GPT-4o model achieves only 44.70% accuracy, well below the 82.02% attained by trained human specialists/experts. We believe that ExAct will be beneficial for developing and evaluating VLMs capable of precise understanding of human skills in various physical and procedural domains. Dataset and code are available at https://texaser.github.io/exact_project_page/

Large language models (LLMs) are increasingly applied to scientific research, yet prevailing science benchmarks probe decontextualized knowledge and overlook the iterative reasoning, hypothesis generation, and observation interpretation that drive scientific discovery. We introduce a scenario-grounded benchmark that evaluates LLMs across biology, chemistry, materials, and physics, where domain experts define research projects of genuine interest and decompose them into modular research scenarios from which vetted questions are sampled. The framework assesses models at two levels: (i) question-level accuracy on scenario-tied items and (ii) project-level performance, where models must propose testable hypotheses, design simulations or experiments, and interpret results. Applying this two-phase scientific discovery evaluation (SDE) framework to state-of-the-art LLMs reveals a consistent performance gap relative to general science benchmarks, diminishing return of scaling up model sizes and reasoning, and systematic weaknesses shared across top-tier models from different providers. Large performance variation in research scenarios leads to changing choices of the best performing model on scientific discovery projects evaluated, suggesting all current LLMs are distant to general scientific "superintelligence". Nevertheless, LLMs already demonstrate promise in a great variety of scientific discovery projects, including cases where constituent scenario scores are low, highlighting the role of guided exploration and serendipity in discovery. This SDE framework offers a reproducible benchmark for discovery-relevant evaluation of LLMs and charts practical paths to advance their development toward scientific discovery.

Assessing student handwritten scratchwork is crucial for personalized educational feedback but presents unique challenges due to diverse handwriting, complex layouts, and varied problem-solving approaches. Existing educational NLP primarily focuses on textual responses and neglects the complexity and multimodality inherent in authentic handwritten scratchwork. Current multimodal large language models (MLLMs) excel at visual reasoning but typically adopt an "examinee perspective", prioritizing generating correct answers rather than diagnosing student errors. To bridge these gaps, we introduce ScratchMath, a novel benchmark specifically designed for explaining and classifying errors in authentic handwritten mathematics scratchwork. Our dataset comprises 1,720 mathematics samples from Chinese primary and middle school students, supporting two key tasks: Error Cause Explanation (ECE) and Error Cause Classification (ECC), with seven defined error types. The dataset is meticulously annotated through rigorous human-machine collaborative approaches involving multiple stages of expert labeling, review, and verification. We systematically evaluate 16 leading MLLMs on ScratchMath, revealing significant performance gaps relative to human experts, especially in visual recognition and logical reasoning. Proprietary models notably outperform open-source models, with large reasoning models showing strong potential for error explanation. All evaluation data and frameworks are publicly available to facilitate further research.

Geometry is a fundamental branch of mathematics and plays a crucial role in evaluating the reasoning capabilities of multimodal large language models (MLLMs). However, existing multimodal mathematics benchmarks mainly focus on plane geometry and largely ignore solid geometry, which requires spatial reasoning and is more challenging than plane geometry. To address this critical gap, we introduce SolidGeo, the first large-scale benchmark specifically designed to evaluate the performance of MLLMs on mathematical reasoning tasks in solid geometry. SolidGeo consists of 3,113 real-world K-12 and competition-level problems, each paired with visual context and annotated with difficulty levels and fine-grained solid geometry categories. Our benchmark covers a wide range of 3D reasoning subjects such as projection, unfolding, spatial measurement, and spatial vector, offering a rigorous testbed for assessing solid geometry. Through extensive experiments, we observe that MLLMs encounter substantial challenges in solid geometry math tasks, with a considerable performance gap relative to human capabilities on SolidGeo. Moreover, we analyze the performance, inference efficiency and error patterns of various models, offering insights into the solid geometric mathematical reasoning capabilities of MLLMs. We hope SolidGeo serves as a catalyst for advancing MLLMs toward deeper geometric reasoning and spatial intelligence.

Multimodal Large Language Models (MLLMs) have shown promising capabilities in mathematical reasoning within visual contexts across various datasets. However, most existing multimodal math benchmarks are limited to single-visual contexts, which diverges from the multi-visual scenarios commonly encountered in real-world mathematical applications. To address this gap, we introduce MV-MATH: a meticulously curated dataset of 2,009 high-quality mathematical problems. Each problem integrates multiple images interleaved with text, derived from authentic K-12 scenarios, and enriched with detailed annotations. MV-MATH includes multiple-choice, free-form, and multi-step questions, covering 11 subject areas across 3 difficulty levels, and serves as a comprehensive and rigorous benchmark for assessing MLLMs' mathematical reasoning in multi-visual contexts. Through extensive experimentation, we observe that MLLMs encounter substantial challenges in multi-visual math tasks, with a considerable performance gap relative to human capabilities on MV-MATH. Furthermore, we analyze the performance and error patterns of various models, providing insights into MLLMs' mathematical reasoning capabilities within multi-visual settings.

Training large models with distributed data parallelism (DDP) requires frequent communication of gradients across workers, which can saturate bandwidth. Infrequent communication strategies (e.g., Local SGD) reduce this overhead but, when applied to adaptive optimizers, often suffer a performance gap relative to fully synchronous DDP. We trace this gap to a time-scale mismatch: the optimizer's fast-moving momentum, tuned for frequent updates, decays too quickly to smooth gradients over long intervals, leading to noise-dominated optimization. To address this, we propose MT-DAO, a family of optimizers that employs multiple slow- and fast-moving first momenta or the gradient to track update dynamics across different time scales, for which we provide the first convergence guarantees. Empirically, for language-model pre-training, this eliminates the performance gap with DDP, outperforming infrequent-communication baselines in perplexity and reducing iso-token wall-clock time by 6-27% on Ethernet interconnects. At the 720M scale, MT-DAO reaches a target perplexity in 24% fewer steps and 35% less time than the single-momentum DDP baseline. MT-DAO enables effective cross-datacenter training and training over wide geographic areas.

We introduce ABot-OCR, an end-to-end vision-language model that transcribes a page image directly into clean Markdown in a single forward pass. By doing so, our approach completely eliminates the need for brittle modular orchestration. To maximize parsing fidelity, we develop a dedicated data engine to provide large-scale, structurally consistent supervision. Furthermore, we propose Decoupled Heterogeneous Document Optimization, a structure-constrained reinforcement learning method that sharpens textual accuracy and strictly enforces markup well-formedness beyond supervised fine-tuning alone. Extensive evaluations demonstrate the superior performance of our framework. On the OmniDocBench v1.5 and v1.6 benchmarks, ABot-OCR achieves state-of-the-art scores of 92.81 and 93.30 among all end-to-end systems, substantially narrowing the performance gap relative to strong pipeline baselines. Finally, comprehensive multilingual text recognition across ten diverse languages further confirms the robust generalizability of ABot-OCR.

Instruction tuned large language models (LLMs), such as ChatGPT, demonstrate remarkable performance in a wide range of tasks. Despite numerous recent studies that examine the performance of instruction-tuned LLMs on various NLP benchmarks, there remains a lack of comprehensive investigation into their ability to understand cross-lingual sociopragmatic meaning (SM), i.e., meaning embedded within social and interactive contexts. This deficiency arises partly from SM not being adequately represented in any of the existing benchmarks. To address this gap, we present SPARROW, an extensive multilingual benchmark specifically designed for SM understanding. SPARROW comprises 169 datasets covering 13 task types across six primary categories (e.g., anti-social language detection, emotion recognition). SPARROW datasets encompass 64 different languages originating from 12 language families representing 16 writing scripts. We evaluate the performance of various multilingual pretrained language models (e.g., mT5) and instruction-tuned LLMs (e.g., BLOOMZ, ChatGPT) on SPARROW through fine-tuning, zero-shot, and/or few-shot learning. Our comprehensive analysis reveals that existing open-source instruction tuned LLMs still struggle to understand SM across various languages, performing close to a random baseline in some cases. We also find that although ChatGPT outperforms many LLMs, it still falls behind task-specific finetuned models with a gap of 12.19 SPARROW score. Our benchmark is available at: https://github.com/UBC-NLP/SPARROW

We present a fine-grained theoretical analysis of the performance gap between reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) under a representation gap. Our study decomposes this gap into two sources: an explicit representation gap under exact optimization and an implicit representation gap under finite samples. In the exact optimization setting, we characterize how the relative capacities of the reward and policy model classes influence the final policy qualities. We show that RLHF, DPO, or online DPO can outperform one another depending on the type of model mis-specifications. Notably, online DPO can outperform both RLHF and standard DPO when the reward and policy model classes are isomorphic and both mis-specified. In the approximate optimization setting, we provide a concrete construction where the ground-truth reward is implicitly sparse and show that RLHF requires significantly fewer samples than DPO to recover an effective reward model -- highlighting a statistical advantage of two-stage learning. Together, these results provide a comprehensive understanding of the performance gap between RLHF and DPO under various settings, and offer practical insights into when each method is preferred.

Although modern multilingual Automatic Speech Recognition (ASR) systems support several Nigerian languages, their performance consistently lags behind high-resource languages like English and French. Nigerian languages present unique modelling hurdles, including acute data scarcity, inconsistent orthography, tonal diacritics, diverse accents, frequent code-switching, and localized named entities. To address these challenges, we developed a multilingual ASR framework utilizing a two-stage distillation process. First, we employ student-teacher knowledge distillation from existing monolingual models, conditioned on robust language-specific N-gram language models. Second, we perform iterative self improvement using pseudo-labelled data to further refine accuracy. Our method significantly bridges the performance gap, achieving on average a relative Word Error Rate (WER) reduction of 29 % over monolingual baselines. Our models also outperform state-of-the-art multilingual models across major benchmarks, including Common Voice and Fleurs. We introduce Sometin Beta Pass Notin (SBPN), a foundational multilingual ASR model covering Yorùbá, Hausa, Igbo, Nigerian Pidgin, and Nigerian English. SBPN is released in two sizes: SBPN-Base (120 M parameters) and SBPN-Large (600 M parameters). By releasing these as open foundation models, we aim to provide ASR resources for further research into the rich phonetic and cultural landscape of the region.

Simulating human profiles by instilling personas into large language models (LLMs) is rapidly transforming research in agentic behavioral simulation, LLM personalization, and human-AI alignment. However, most existing synthetic personas remain shallow and simplistic, capturing minimal attributes and failing to reflect the rich complexity and diversity of real human identities. We introduce DEEPPERSONA, a scalable generative engine for synthesizing narrative-complete synthetic personas through a two-stage, taxonomy-guided method. First, we algorithmically construct the largest-ever human-attribute taxonomy, comprising over hundreds of hierarchically organized attributes, by mining thousands of real user-ChatGPT conversations. Second, we progressively sample attributes from this taxonomy, conditionally generating coherent and realistic personas that average hundreds of structured attributes and roughly 1 MB of narrative text, two orders of magnitude deeper than prior works. Intrinsic evaluations confirm significant improvements in attribute diversity (32 percent higher coverage) and profile uniqueness (44 percent greater) compared to state-of-the-art baselines. Extrinsically, our personas enhance GPT-4.1-mini's personalized question answering accuracy by 11.6 percent on average across ten metrics and substantially narrow (by 31.7 percent) the gap between simulated LLM citizens and authentic human responses in social surveys. Our generated national citizens reduced the performance gap on the Big Five personality test by 17 percent relative to LLM-simulated citizens. DEEPPERSONA thus provides a rigorous, scalable, and privacy-free platform for high-fidelity human simulation and personalized AI research.

Practical object detection application can lose its effectiveness on image inputs with natural distribution shifts. This problem leads the research community to pay more attention on the robustness of detectors under Out-Of-Distribution (OOD) inputs. Existing works construct datasets to benchmark the detector's OOD robustness for a specific application scenario, e.g., Autonomous Driving. However, these datasets lack universality and are hard to benchmark general detectors built on common tasks such as COCO. To give a more comprehensive robustness assessment, we introduce COCO-O(ut-of-distribution), a test dataset based on COCO with 6 types of natural distribution shifts. COCO-O has a large distribution gap with training data and results in a significant 55.7% relative performance drop on a Faster R-CNN detector. We leverage COCO-O to conduct experiments on more than 100 modern object detectors to investigate if their improvements are credible or just over-fitting to the COCO test set. Unfortunately, most classic detectors in early years do not exhibit strong OOD generalization. We further study the robustness effect on recent breakthroughs of detector's architecture design, augmentation and pre-training techniques. Some empirical findings are revealed: 1) Compared with detection head or neck, backbone is the most important part for robustness; 2) An end-to-end detection transformer design brings no enhancement, and may even reduce robustness; 3) Large-scale foundation models have made a great leap on robust object detection. We hope our COCO-O could provide a rich testbed for robustness study of object detection. The dataset will be available at https://github.com/alibaba/easyrobust/tree/main/benchmarks/coco_o.

Sparse attention can effectively mitigate the significant memory and throughput demands of Large Language Models (LLMs) in long contexts. Existing methods typically employ a uniform sparse attention mask, applying the same sparse pattern across different attention heads and input lengths. However, this uniform approach fails to capture the diverse attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose the Mixture of Attention (MoA), which automatically tailors distinct sparse attention configurations to different heads and layers. MoA constructs and navigates a search space of various attention patterns and their scaling rules relative to input sequence lengths. It profiles the model, evaluates potential configurations, and pinpoints the optimal sparse attention compression plan. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer sequences, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by 3.9times with the same average attention span, boosting retrieval accuracy by 1.5-7.1times over the uniform-attention baseline across Vicuna-7B, Vicuna-13B, and Llama3-8B models. Moreover, MoA narrows the capability gaps between sparse and dense models, reducing the maximum relative performance drop from 9%-36% to within 5% across two long-context understanding benchmarks. MoA achieves a 1.2-1.4times GPU memory reduction and boosts decode throughput by 5.5-6.7 times for 7B and 13B dense models on a single GPU, with minimal impact on performance.

Generative spoken language models pretrained on large-scale raw audio can continue a speech prompt with appropriate content while preserving attributes like speaker and emotion, serving as foundation models for spoken dialogue. In prior literature, these models are often evaluated using ``global token perplexity'', which directly applies the text perplexity formulation to speech tokens. However, this practice overlooks fundamental differences between speech and text modalities, possibly leading to an underestimation of the speech characteristics. In this work, we propose a variety of likelihood- and generative-based evaluation methods that serve in place of naive global token perplexity. We demonstrate that the proposed evaluations more faithfully reflect perceived generation quality, as evidenced by stronger correlations with human-rated mean opinion scores (MOS). When assessed under the new metrics, the relative performance landscape of spoken language models is reshaped, revealing a significantly reduced gap between the best-performing model and the human topline. Together, these results suggest that appropriate evaluation is critical for accurately assessing progress in spoken language modeling.

Document Layout analysis (DLA), is the process by which a page is parsed into meaningful elements, often using machine learning models. Typically, the quality of a model is judged using general object detection metrics such as IoU, F1 or mAP. However, these metrics are designed for images that are 2D projections of 3D space, not for the natively 2D imagery of printed media. This discrepancy can result in misleading or uninformative interpretation of model performance by the metrics. To encourage more robust, comparable, and nuanced DLA, we introduce: The Structural Semantic Unit (SSU) a relational labelling approach that shifts the focus from the physical to the semantic structure of the content; and the Coverage, Overlap, Trespass, and Excess (COTe) score, a decomposable metric for measuring page parsing quality. We demonstrate the value of these methods through case studies and by evaluating 5 common DLA models on 3 DLA datasets. We show that the COTe score is more informative than traditional metrics and reveals distinct failure modes across models, such as breaching semantic boundaries or repeatedly parsing the same region. In addition, the COTe score reduces the interpretation-performance gap by up to 76% relative to the F1. Notably, we find that the COTe's granularity robustness largely holds even without explicit SSU labelling, lowering the barriers to entry for using the system. Finally, we release an SSU labelled dataset and a Python library for applying COTe in DLA projects.

We introduce Complex-Edit, a comprehensive benchmark designed to systematically evaluate instruction-based image editing models across instructions of varying complexity. To develop this benchmark, we harness GPT-4o to automatically collect a diverse set of editing instructions at scale. Our approach follows a well-structured ``Chain-of-Edit'' pipeline: we first generate individual atomic editing tasks independently and then integrate them to form cohesive, complex instructions. Additionally, we introduce a suite of metrics to assess various aspects of editing performance, along with a VLM-based auto-evaluation pipeline that supports large-scale assessments. Our benchmark yields several notable insights: 1) Open-source models significantly underperform relative to proprietary, closed-source models, with the performance gap widening as instruction complexity increases; 2) Increased instructional complexity primarily impairs the models' ability to retain key elements from the input images and to preserve the overall aesthetic quality; 3) Decomposing a complex instruction into a sequence of atomic steps, executed in a step-by-step manner, substantially degrades performance across multiple metrics; 4) A straightforward Best-of-N selection strategy improves results for both direct editing and the step-by-step sequential approach; and 5) We observe a ``curse of synthetic data'': when synthetic data is involved in model training, the edited images from such models tend to appear increasingly synthetic as the complexity of the editing instructions rises -- a phenomenon that intriguingly also manifests in the latest GPT-4o outputs.

Large language models (LLMs) show increasing promise in medical applications, but their ability to detect and correct errors in clinical texts -- a prerequisite for safe deployment -- remains under-evaluated, particularly beyond English. We introduce MedRECT, a cross-lingual benchmark (Japanese/English) that formulates medical error handling as three subtasks: error detection, error localization (sentence extraction), and error correction. MedRECT is built with a scalable, automated pipeline from the Japanese Medical Licensing Examinations (JMLE) and a curated English counterpart, yielding MedRECT-ja (663 texts) and MedRECT-en (458 texts) with comparable error/no-error balance. We evaluate 9 contemporary LLMs spanning proprietary, open-weight, and reasoning families. Key findings: (i) reasoning models substantially outperform standard architectures, with up to 13.5% relative improvement in error detection and 51.0% in sentence extraction; (ii) cross-lingual evaluation reveals 5-10% performance gaps from English to Japanese, with smaller disparities for reasoning models; (iii) targeted LoRA fine-tuning yields asymmetric improvements in error correction performance (Japanese: +0.078, English: +0.168) while preserving reasoning capabilities; and (iv) our fine-tuned model exceeds human expert performance on structured medical error correction tasks. To our knowledge, MedRECT is the first comprehensive cross-lingual benchmark for medical error correction, providing a reproducible framework and resources for developing safer medical LLMs across languages.

Memory-efficient optimization is critical for training increasingly large language models (LLMs). A popular strategy involves gradient low-rank projection, storing only the projected optimizer states, with GaLore being a representative example. However, a significant drawback of many such methods is their lack of convergence guarantees, as various low-rank projection approaches introduce inherent biases relative to the original optimization algorithms, which contribute to performance gaps compared to full-parameter training. Aiming to tackle this problem, this paper investigates the layerwise sampling technique for debiasing low-rank projection mechanisms. In particular, an instantiation of the paradigm gives rise to a novel and unbiased low-rank optimization method built upon GaLore's mechanism and the Muon algorithm, named GaLore Unbiased with Muon (GUM). We theoretically prove our method matches the convergence guarantees of the base Muon algorithm while preserving the memory efficiency of low-rank techniques. Empirical experiments on LLM fine-tuning and pretraining also demonstrate non-trivial improvements over GaLore and even better performance than full-parameter training. Further investigation shows that the improvement of this technique comes from a more uniform distribution of knowledge inside layers, leading to more efficient utilization of the model parameter space and better memorization.

Recent reasoning Large Language Models produce a chain-of-thought (CoT) predominantly in English, even when prompted in non-English languages. Prior work suggests that forcing the CoT to remain in the input language (native reasoning) substantially degrades performance relative to allowing the model to reason in English before answering in the input language (English-pivoted reasoning). However, most studies of this native reasoning gap rely on inference-time interventions or limited native-language training data. We revisit this comparison at a larger scale and under comparable supervision. We construct long multilingual reasoning datasets across six languages (English, French, German, Spanish, Chinese and Swahili); fine-tune specialists in both native and English-pivoted regimes on top of Qwen/Qwen3-8B-Base, and evaluate across mathematics, science, general knowledge, and code. In this setting, the average native reasoning gap shrinks to 1.9--3.5\% across the five non-English languages, considerably smaller than previously reported. Weight-space analysis of the native specialists reveals aligned fine-tuning updates in the middle layers and divergence in the outer layers. This points to a largely language-agnostic reasoning core surrounded by language-specific layers. Exploiting this structure, we introduce a Layer Swap: transferring the English specialist's stronger reasoning mid-layers into each native specialist, closing most of the native reasoning gap across the five non-English languages while preserving CoT in the target language. We release all models and datasets.

Long-horizon agents need a context manager to compress growing interaction histories into a bounded working context, via passive strategies or active strategies that decide how memory is accessed and reorganized. Meanwhile, prior optical-memory work mainly treats pixels as a dense codec for textualized histories, often presupposing that rendering context into optical memory incurs a significant performance drop relative to text, thus coupling this representation with SFT, self-distillation, or reinforcement learning to close this gap, leaving unresolved (i) how visual rendering performs as a context manager under a fair, controlled comparison, and (ii) whether this carrier offers a native advantage when history is inherently multimodal. In this paper, we formulate context management as a budget-constrained history transformation and introduce Visual Rendering (VR) as a representational context manager. Under a shared harness, policy model, trigger, and task domain, we evaluate VR on four text-centric and three multimodal benchmarks against four baselines (No Compression, Discard-All, Sliding Window, Summarization), finding visual memory is a natural carrier of native visual evidence. Building on this finding, we propose VERA (Visual Evidence-Retaining strategy for long-horizon Agents), a training-free context manager built on deterministic rendering with no exposed memory operations: on text-centric benchmarks it renders textual history as VR does, while on multimodal benchmarks it retains native visual observations instead of translating them into text. Across nearly all benchmarks, VERA cuts cumulative non-cache tokens by 31.5%-63.1% versus No Compression, matches existing managers on text-centric tasks, and achieves the highest accuracy among all baselines on multimodal tasks, supporting a modality-preserving view of long-horizon context management.

Current Text-to-SQL methods are evaluated and only focused on executable queries, overlooking the semantic alignment challenge -- both in terms of the semantic meaning of the query and the correctness of the execution results. Even execution accuracy itself shows significant drops when moving from English to other languages, with an average decline of 6 percentage points across non-English languages. We address these challenges by presenting a new framework that combines Group Relative Policy Optimization (GRPO) within a multilingual contrastive reward signal to enhance both task efficiency and semantic accuracy in Text-to-SQL systems in cross-lingual scenarios. Our method teaches models to obtain better correspondence between SQL generation and user intent by combining a reward signal based on semantic similarity. On the seven-language MultiSpider dataset, fine-tuning the LLaMA-3-3B model with GRPO improved the execution accuracy up to 87.4 percent (+26 pp over zero-shot) and semantic accuracy up to 52.29 percent (+32.86 pp). Adding our contrastive reward signal in the GRPO framework further improved the average semantic accuracy to 59.14 percent (+6.85 pp, up to +10 pp for Vietnamese). Our experiments showcase that a smaller, parameter-efficient 3B LLaMA model fine-tuned with our contrastive reward signal outperforms a much larger zero-shot 8B LLaMA model, with an uplift of 7.43 pp in execution accuracy (from 81.43 percent on the 8B model to 88.86 percent on the 3B model), and nearly matches its semantic accuracy (59.14 percent vs. 68.57 percent) -- all using just 3,000 reinforcement learning training examples. These results demonstrate how we can improve the performance of Text-to-SQL systems with contrastive rewards for directed semantic alignment, without requiring large-scale training datasets.

Reinforcement learning (RL) is widely used for post-training large language models (LLMs) in code editing, where group-relative methods, such as GRPO, are popular due to their critic-free and normalized advantage estimation. However, in real-world code-editing scenarios, reward distributions are often skewed with unpredictable noise, leading to distorted advantage computation and increased rollout outliers. To address this issue, we propose Group Adaptive Policy Optimization (GAPO), which adaptively finds an interval with the highest SNR (Signal to Noise Ratio) per prompt and uses the median of that interval as an adaptive Q to replace the group mean in advantage calculation to reduce noise further. This adaptive Q robustly handles rollout noise while remaining plug-and-play and efficient. We evaluate GAPO on nine instruction-tuned LLMs (3B-14B) using a collected large dataset of 51,844 real-world, history-aware code-editing tasks spanning 10 programming languages. GAPO yields up to 4.35 in-domain (ID) and 5.30 out-of-domain (OOD) exact-match improvements over GRPO and its variant DAPO, while achieving lower clipping ratios and higher GPU throughput. Code: https://github.com/TsingZ0/verl-GAPO.

Multi-teacher on-policy distillation (MOPD) is used in two settings. In common-domain composition, several teachers score each student rollout from one prompt domain and their signals form a single target; in routed-domain distillation, prompts from different domains are assigned to the corresponding specialist. Both settings usually transfer each teacher's endpoint policy, which mixes what post-training changed with preferences inherited from the teacher's base. We introduce Δ-MOPD, which transfers each teacher's teacher-minus-base logit shift re-anchored at the student's frozen initialization, and compare it with endpoint supervision in both settings while holding teacher selection fixed. We first expose the mechanism that impedes endpoint transfer: inherited base pull can exceed the post-training shift. Removing it reduces the teacher-term norm ratio and target--student KL. Across our experiments, the results suggest that shift targets are particularly useful when teacher signals are combined at a state. With three composed teachers, Δ-MOPD exceeds endpoint composition by 4.11 Math and 1.95 five-benchmark points; with two, it matches endpoint accuracy. Under phased routing, it achieves higher mean performance in both phase orders and reduces the observed order gap from 10.50 to 6.42 points. Under interleaved routing, where each update involves one teacher, the two targets perform comparably. The phased results provide supporting evidence that the benefit may extend to signals accumulated across training phases. Target construction is thus an independent design axis in MOPD, complementary to teacher selection.

Self-play fine-tuning has demonstrated promising abilities in adapting large language models (LLMs) to downstream tasks with limited real-world data. The basic principle is to iteratively refine the model with real samples and synthetic ones generated from itself. However, the existing methods primarily focus on the relative gaps between the rewards for two types of data, neglecting their absolute values. Through theoretical analysis, we identify that the gap-based methods suffer from unstable evolution, due to the potentially degenerated objectives. To address this limitation, we introduce a novel self-play fine-tuning method, namely Self-PlAy via Noise Contrastive Estimation (SPACE), which leverages noise contrastive estimation to capture the real-world data distribution. Specifically, SPACE treats synthetic samples as auxiliary components, and discriminates them from the real ones in a binary classification manner. As a result, SPACE independently optimizes the absolute reward values for each type of data, ensuring a consistently meaningful objective and thereby avoiding the instability issue. Theoretically, we show that the optimal solution of the objective in SPACE aligns with the underlying distribution of real-world data, and SPACE guarantees a provably stable convergence to the optimal distribution. Empirically, we show that SPACE significantly improves the performance of LLMs over various tasks, and outperforms supervised fine-tuning that employs much more real-world samples. Compared to gap-based self-play fine-tuning methods, SPACE exhibits remarkable superiority and stable evolution.

With VLM-powered computer-using agents (CUAs) becoming increasingly capable at graphical user interface (GUI) navigation and manipulation, reliable step-level decision-making has emerged as a key bottleneck for real-world deployment. In long-horizon workflows, errors accumulate quickly and irreversible actions can cause unintended consequences, motivating critic models that assess each action before execution. While critic models offer a promising solution, their effectiveness is hindered by the lack of diverse, high-quality GUI feedback data and public critic benchmarks for step-level evaluation in computer use. To bridge these gaps, we introduce OS-Oracle that makes three core contributions: (1) a scalable data pipeline for synthesizing cross-platform GUI critic data; (2) a two-stage training paradigm combining supervised fine-tuning (SFT) and consistency-preserving group relative policy optimization (CP-GRPO); (3) OS-Critic Bench, a holistic benchmark for evaluating critic model performance across Mobile, Web, and Desktop platforms. Leveraging this framework, we curate a high-quality dataset containing 310k critic samples. The resulting critic model, OS-Oracle-7B, achieves state-of-the-art performance among open-source VLMs on OS-Critic Bench, and surpasses proprietary models on the mobile domain. Furthermore, when serving as a pre-critic, OS-Oracle-7B improves the performance of native GUI agents such as UI-TARS-1.5-7B in OSWorld and AndroidWorld environments. The code is open-sourced at https://github.com/numbmelon/OS-Oracle.

Current Vision-Language Models (VLMs) are celebrated for their multilingual capabilities, yet they operate under a flawed assumption: that one language corresponds to a single writing system. This overlooks billions of users of multi-script languages like Punjabi, Serbian, Hindi-Urdu, Kurdish, among many others, for whom a model's capability may be fractured by orthographic bias. We introduce PuMVR (Punjabi Multimodal Visual Reasoning), the first benchmark designed to quantify script-dependent bias through 375 culturally grounded image-reasoning tasks across Punjabi's three active scripts (Gurmukhi, Shahmukhi, Roman). Evaluating 10 state-of-the-art VLMs, we expose a substantial Script Gap: models frequently solve visual puzzles in one script while failing identical tasks in another, with accuracy deltas reaching 16% and Script Consistency Rates (SCR) as low as 24.8%. Crucially, visual input boosts absolute performance but does not close this gap, the relative bias persists. Our analysis suggests reasoning patterns show limited cross-script transferability, and Chain-of-Thought pathways diverge based on script alone. We propose SCR as a core metric for script-agnostic evaluation, challenging current multilingual assessment paradigms and providing a framework for equitable AI.

Embodied Vision-Language-Action (VLA) models are typically obtained by fine-tuning powerful pretrained VLMs on robotics data, yet it is unclear how much commonsense and factual knowledge they retain after adaptation. Failures on knowledge-sensitive tasks are ambiguous, conflating missing knowledge with poor generalization of low-level control. We introduce Act2Answer, a lightweight protocol that adapts VLM knowledge benchmarks to VLA evaluation by requiring agents to answer through action. Each question becomes a short tabletop episode where the agent performs a single object-placement action to select among candidate answers, yielding an action-grounded success rate with reduced control confounds. We curate a test suite of such environments across diverse commonsense and world-knowledge categories and introduce layerwise intent probing to localize answer-relevant information across the VLM backbone and action head. In a large-scale study of 7 VLA models and 9 VLM baselines, we systematically rank models across categories, finding that VLAs show solid performance on simple concepts while exhibiting larger gaps on richer semantic categories relative to their source VLMs, that VQA co-training is associated with better knowledge retention, and that answer-relevant signals peak in middle VLA layers but attenuate in upper layers. Act2Answer is available at https://tttonyalpha.github.io/act2answer/.

Large Multimodal Models (LMMs) large-scale deployment in industrial warehouse settings specifically necessitates that models exhibit human-expert-level hazard-oriented perception, understanding, and reasoning capabilities. However, the scarcity of real industrial data, tightly coupled to commercial terms, significantly hampers further advancement. To bridge this gap, we curate LogiScope-VQA to investigate the practical applicability of mainstream LMMs in real-world logistics operations. LogiScope-VQA comprises 2,476 images and 2,918 videos primarily sourced from real-world logistics parks, along with 10,274 VQAs meticulously curated and validated by human annotators. Grounded in 18 core objects and 20 risk types, we devise 39 subtasks aligned with three principal themes: industrial element perception, warehouse knowledge understanding, and potential risk reasoning. Furthermore, we incorporate dynamic thinking-budget configurations and dual-dimensional risk bias analyses to elucidate the properties of LMMs. Extensive experiments unveil that even powerful proprietary models, including GPT-5.5, Gemini-3.1-Pro, and Claude-Opus-4.7, exhibit a significant gap relative to human performance. The unique challenge of jointly integrating perception, understanding, and reasoning for hazard identification poses substantial headroom for further improvement on LogiScope-VQA. We additionally reveal the pervasive security bias issue that impedes LLMs' practical deployment in real-world settings. The industrial dataset is publicly available under the CC BY-NC-SA 4.0 license.

AI agents operating on user interfaces must understand how interfaces communicate state and feedback to act reliably. As a core communicative modality, animations are increasingly used in modern interfaces, serving critical functional purposes beyond mere aesthetics. Thus, understanding UI animation is essential for comprehensive interface interpretation. However, recent studies of Vision Language Models (VLMs) for UI understanding have focused primarily on static screenshots, leaving it unclear how well these models handle dynamic UI animations. To address this gap, we created AniMINT, a novel dataset of 300 densely annotated UI animation videos. We systematically evaluate state-of-the-art VLMs on UI animation understanding, including their abilities to perceive the animation effects, identify animation purposes, and interpret animation meaning. Our results show that VLMs can reliably detect primitive motion. However, their high-level animation interpretation remains inconsistent, with substantial gaps relative to human performance. Finally, we use Motion, Context, and Perceptual Cues (MCPC) to probe factors affecting VLM performance, revealing key bottlenecks and directions for future improvement.

Monocular and binocular self-supervised depth estimations are two important and related tasks in computer vision, which aim to predict scene depths from single images and stereo image pairs respectively. In literature, the two tasks are usually tackled separately by two different kinds of models, and binocular models generally fail to predict depth from single images, while the prediction accuracy of monocular models is generally inferior to binocular models. In this paper, we propose a Two-in-One self-supervised depth estimation network, called TiO-Depth, which could not only compatibly handle the two tasks, but also improve the prediction accuracy. TiO-Depth employs a Siamese architecture and each sub-network of it could be used as a monocular depth estimation model. For binocular depth estimation, a Monocular Feature Matching module is proposed for incorporating the stereo knowledge between the two images, and the full TiO-Depth is used to predict depths. We also design a multi-stage joint-training strategy for improving the performances of TiO-Depth in both two tasks by combining the relative advantages of them. Experimental results on the KITTI, Cityscapes, and DDAD datasets demonstrate that TiO-Depth outperforms both the monocular and binocular state-of-the-art methods in most cases, and further verify the feasibility of a two-in-one network for monocular and binocular depth estimation. The code is available at https://github.com/ZM-Zhou/TiO-Depth_pytorch.

Agentic search -- the task of training agents that iteratively reason, issue queries, and synthesize retrieved information to answer complex questions -- has achieved remarkable progress through reinforcement learning (RL). However, existing approaches such as Search-R1, treat the retrieval system as a fixed tool, optimizing only the reasoning agent while the retrieval component remains unchanged. A preliminary experiment reveals that the gap between an oracle and a fixed retrieval system reaches up to +26.8% relative F1 improvement across seven QA benchmarks, suggesting that the retrieval system is a key bottleneck in scaling agentic search performance. Motivated by this finding, we propose CoSearch, a framework that jointly trains a multi-step reasoning agent and a generative document ranking model via Group Relative Policy Optimization (GRPO). To enable effective GRPO training for the ranker -- whose inputs vary across reasoning trajectories -- we introduce a semantic grouping strategy that clusters sub-queries by token-level similarity, forming valid optimization groups without additional rollouts. We further design a composite reward combining ranking quality signals with trajectory-level outcome feedback, providing the ranker with both immediate and long-term learning signals. Experiments on seven single-hop and multi-hop QA benchmarks demonstrate consistent improvements over strong baselines, with ablation studies validating each design choice. Our results show that joint training of the reasoning agent and retrieval system is both feasible and strongly performant, pointing to a key ingredient for future search agents.

Leaderboards rank long-horizon agents by their final outputs. Yet a higher score alone does not establish whether two systems are comparable or which stage accounts for the difference. Unequal evidence, inputs, or budgets can affect scores, and statistical corrections do not remove this mismatch. We introduce compositional controllability to address these questions. A comparison window covers one stage, several stages, or the whole agent. Our central result bounds the gap between observed and controlled score differences using only nuisance outside the window. This yields an admissibility test applied before scores are inspected. Inadmissible comparisons are refused. For admissible pairs, an ordering is certified only when the score gap exceeds the combined sampling and nuisance radii; otherwise, it remains undecided. These decisions give each system a rank interval. We introduce BioLitBench, a benchmark of 2,042 biomedical articles represented as structured claim graphs. Among seven published pipelines, a conventional statistical analysis declares a winner in 14 of 21 pairwise comparisons. Yet the top-ranked system alone received the target review's bibliography. To isolate pipeline performance, our test requires matched inputs and a fixed backbone model. It refuses 11 of the 21 comparisons, including every comparison involving the top-ranked system. Seven of the 14 conventional conclusions fall within these refused pairs. The same comparison windows support stage-level training. We train SCRIBE on Qwen3.8-27B using rewards measured at each stage's exit. Under matched evidence, SCRIBE achieves a certified rank interval of [1,2], with certified advantages over all evaluated published pipelines and the evaluated Claude and OpenAI agents. Under same pool, SCRIBE matches the strongest published retriever and is certified above three published pipelines.

Majority voting over multiple LLM attempts improves mathematical reasoning, but correlated errors limit the effective sample size. A natural fix is to assign different reasoning strategies to different voters. The approach, Diverse Prompt Mixer, is tested on the AIMO 3 competition: 3 models, 23+ experiments, 50 IMO-level problems, one H100 80 GB, 5-hour limit. Every prompt-level intervention fails. High-temperature sampling already decorrelates errors; weaker strategies reduce accuracy more than they reduce correlation. Across an 8-point capability gap at equal N=8 and every optimization tested, model capability dominates. The gap between the best majority-vote score (42/50) and pass@20 (~45.5) is selection loss, not prompt loss. A verifier-based selector could close it. Prompt engineering cannot.

In multi-task learning (MTL), gradient balancing has recently attracted more research interest than loss balancing since it often leads to better performance. However, loss balancing is much more efficient than gradient balancing, and thus it is still worth further exploration in MTL. Note that prior studies typically ignore that there exist varying improvable gaps across multiple tasks, where the improvable gap per task is defined as the distance between the current training progress and desired final training progress. Therefore, after loss balancing, the performance imbalance still arises in many cases. In this paper, following the loss balancing framework, we propose two novel improvable gap balancing (IGB) algorithms for MTL: one takes a simple heuristic, and the other (for the first time) deploys deep reinforcement learning for MTL. Particularly, instead of directly balancing the losses in MTL, both algorithms choose to dynamically assign task weights for improvable gap balancing. Moreover, we combine IGB and gradient balancing to show the complementarity between the two types of algorithms. Extensive experiments on two benchmark datasets demonstrate that our IGB algorithms lead to the best results in MTL via loss balancing and achieve further improvements when combined with gradient balancing. Code is available at https://github.com/YanqiDai/IGB4MTL.

The performance of Automatic Speech Recognition (ASR) systems has constantly increased in state-of-the-art development. However, performance tends to decrease considerably in more challenging conditions (e.g., background noise, multiple speaker social conversations) and with more atypical speakers (e.g., children, non-native speakers or people with speech disorders), which signifies that general improvements do not necessarily transfer to applications that rely on ASR, e.g., educational software for younger students or language learners. In this study, we focus on the gap in performance between recognition results for native and non-native, read and spontaneous, Swedish utterances transcribed by different ASR services. We compare the recognition results using Word Error Rate and analyze the linguistic factors that may generate the observed transcription errors.

In many real-world applications, deep neural networks are retrained from scratch as a dataset grows in size. Given the computational expense for retraining networks, it has been argued that continual learning could make updating networks more efficient. An obstacle to achieving this goal is the stability gap, which refers to an observation that when updating on new data, performance on previously learned data degrades before recovering. Addressing this problem would enable learning new data with fewer network updates, resulting in increased computational efficiency. We study how to mitigate the stability gap. We test a variety of hypotheses to understand why the stability gap occurs. This leads us to discover a method that vastly reduces this gap. In large-scale class incremental learning experiments, we are able to significantly reduce the number of network updates needed for continual learning. Our work has the potential to advance the state-of-the-art in continual learning for real-world applications along with reducing the carbon footprint required to maintain updated neural networks.

Vision-Language Models (VLMs) perform well on multimodal benchmarks but lag behind humans and specialized models on visual perception tasks like depth estimation or object counting. Finetuning on one task can unpredictably affect performance on others, making task-specific finetuning challenging. In this paper, we address this challenge through a systematic study of task transferability. We examine how finetuning a VLM on one perception task affects its zero-shot performance on others. We introduce Perfection Gap Factor (PGF), a normalized metric that measures change in performance as a result of task transfer. We utilize PGF to compute Task Transferability, which captures both the breadth and the magnitude of transfer induced by a source task. Using three open-weight VLMs evaluated across 13 perception tasks, we construct a task transfer graph that reveals previously unobserved relationships among perception tasks. Our analysis uncovers patterns of positive and negative transfer, identifies groups of tasks that mutually influence each other, organizes tasks into personas based on their transfer behavior and demonstrates how PGF can guide data selection for more efficient training. These findings highlight both opportunities for positive transfer and risks of negative interference, offering actionable guidance for advancing VLMs.

Large language models cannot estimate how long their own tasks take. We investigate this limitation through four experiments across 68 tasks and four model families. Pre-task estimates overshoot actual duration by 4--7times (p < 0.001), with models predicting human-scale minutes for tasks completing in seconds. Relative ordering fares no better: on task pairs designed to expose heuristic reliance, models score at or below chance (GPT-5: 18\% on counter-intuitive pairs, p = 0.033), systematically failing when complexity labels mislead. Post-hoc recall is disconnected from reality -- estimates diverge from actuals by an order of magnitude in either direction. These failures persist in multi-step agentic settings, with errors of 5--10times. The models possess propositional knowledge about duration from training but lack experiential grounding in their own inference time, with practical implications for agent scheduling, planning and time-critical scenarios.

AI agents increasingly assist with financial research, yet no benchmark evaluates their ability to retrieve specific numeric values from structured databases. We introduce FinRetrieval, a benchmark of 500 financial retrieval questions with ground truth answers, agent responses from 14 configurations across three frontier providers (Anthropic, OpenAI, Google), and complete tool call execution traces. Our evaluation reveals that tool availability dominates performance: Claude Opus achieves 90.8% accuracy with structured data APIs but only 19.8% with web search alone--a 71 percentage point gap that exceeds other providers by 3-4x. We find that reasoning mode benefits vary inversely with base capability (+9.0pp for OpenAI vs +2.8pp for Claude), explained by differences in base-mode tool utilization rather than reasoning ability. Geographic performance gaps (5.6pp US advantage) stem from fiscal year naming conventions, not model limitations. We release the dataset, evaluation code, and tool traces to enable research on financial AI systems.

Vision-language models (VLMs) generate fluent causal explanations, but current evaluations cannot distinguish linguistic plausibility from faithful causal reasoning. We introduce a dual-probe methodology that isolates these properties. The Text-Only Probe measures linguistic quality. The Chain-Text Probe requires models to first generate explicit causal chains. The Abstraction Gap (AG) metric quantifies the normalized performance difference. Evaluating eight VLMs on CAGE (Causal Abstraction Gap Evaluation), a benchmark of 49,500 questions across 5,500 images spanning Pearl's causal hierarchy, we find seven models exhibit AG exceeding 0.50 with text scores of 6--8 but chain scores below 2.5. Fine-tuning on 45,000 chain-annotated examples fails to close the gap. However, one model achieves near-zero AG. The capability exists within current VLM architectures and depends on pretraining and architectural choices. CAGE provides a diagnostic tool for assessing faithful causal reasoning in VLMs.

Background: Deep learning models are typically trained using stochastic gradient descent or one of its variants. These methods update the weights using their gradient, estimated from a small fraction of the training data. It has been observed that when using large batch sizes there is a persistent degradation in generalization performance - known as the "generalization gap" phenomena. Identifying the origin of this gap and closing it had remained an open problem. Contributions: We examine the initial high learning rate training phase. We find that the weight distance from its initialization grows logarithmically with the number of weight updates. We therefore propose a "random walk on random landscape" statistical model which is known to exhibit similar "ultra-slow" diffusion behavior. Following this hypothesis we conducted experiments to show empirically that the "generalization gap" stems from the relatively small number of updates rather than the batch size, and can be completely eliminated by adapting the training regime used. We further investigate different techniques to train models in the large-batch regime and present a novel algorithm named "Ghost Batch Normalization" which enables significant decrease in the generalization gap without increasing the number of updates. To validate our findings we conduct several additional experiments on MNIST, CIFAR-10, CIFAR-100 and ImageNet. Finally, we reassess common practices and beliefs concerning training of deep models and suggest they may not be optimal to achieve good generalization.

Despite the widespread adoption of Large language models (LLMs), their remarkable capabilities remain limited to a few high-resource languages. Additionally, many low-resource languages (e.g. African languages) are often evaluated only on basic text classification tasks due to the lack of appropriate or comprehensive benchmarks outside of high-resource languages. In this paper, we introduce IrokoBench -- a human-translated benchmark dataset for 16 typologically-diverse low-resource African languages covering three tasks: natural language inference~(AfriXNLI), mathematical reasoning~(AfriMGSM), and multi-choice knowledge-based QA~(AfriMMLU). We use IrokoBench to evaluate zero-shot, few-shot, and translate-test settings~(where test sets are translated into English) across 10 open and four proprietary LLMs. Our evaluation reveals a significant performance gap between high-resource languages~(such as English and French) and low-resource African languages. We observe a significant performance gap between open and proprietary models, with the highest performing open model, Aya-101 only at 58\% of the best-performing proprietary model GPT-4o performance. Machine translating the test set to English before evaluation helped to close the gap for larger models that are English-centric, like LLaMa 3 70B. These findings suggest that more efforts are needed to develop and adapt LLMs for African languages.

People overthink; language models over-sample, and the extra effort can talk both into a worse answer. Reasoning systems answer a hard question by sampling it many times (test-time scaling), and the more they draw, the more often a correct answer turns up somewhere, so coverage, the fraction of problems with at least one correct try, climbs and appears to be progress. But a deployed system must return one answer, and choosing it, not knowing which try is right, is selection; selection is capped, and past a point extra samples only make the model surer of a confident mistake, even as every draw adds cost. The gap between climbing coverage and stalled selection, the identifiability gap, is the answer a model can produce but not pick. So the real question is not whether to sample but how far, and the answer is: not far. For picking an answer, the vote has already settled within a few dozen draws, the modal ceiling; for scoring a benchmark, sooner still, the correlation ceiling. Beyond that, extra draws cost compute and add nothing, and can even make the answer worse. This paper turns the cutoff into a single number, the effective number of samples, that any sampling run already reveals. The bottleneck is recognizing a right answer, not generating one.

Code models are increasingly trained with execution feedback, but most training signals still stop at correctness. This leaves an important gap for systems code: two programs can pass the same tests while differing greatly in runtime. We study how to train code agents to prefer faster correct implementations, rather than treating efficiency only as an evaluation metric. The key difficulty is that runtime is a fragile reward. It is meaningful only after a program is correct, varies across tasks, and gives little guidance when most sampled programs fail to compile or run. We propose RLPF, reinforcement learning from performance feedback, which turns execution outcomes into a staged reward. Failed programs are ordered by execution progress, while correct programs are ranked by their relative improvement from the baseline toward the expert reference. This gives useful feedback before correctness and performance-sensitive feedback after correctness. Fine-tuning Qwen3-32B with RLPF on PerfCodeBench raises correct-and-runnable solutions from 11.1% to 54.6% and improves relative efficiency from 8.1% to 38.6%. The trained model becomes competitive with stronger open-weight systems, and its optimization behavior transfers modestly to EffiBench-X. Additional studies show that model-generated references provide useful but weaker supervision, and that the full composite reward is more reliable than correctness-only or runtime-only baselines. These results suggest that code agents can be trained not only to pass tests, but also to optimize the programs they write.

In domains ranging from computer vision to natural language processing, machine learning models have been shown to exhibit stark disparities, often performing worse for members of traditionally underserved groups. One factor contributing to these performance gaps is a lack of representation in the data the models are trained on. It is often unclear, however, how to operationalize representativeness in specific applications. Here we formalize the problem of creating equitable training datasets, and propose a statistical framework for addressing this problem. We consider a setting where a model builder must decide how to allocate a fixed data collection budget to gather training data from different subgroups. We then frame dataset creation as a constrained optimization problem, in which one maximizes a function of group-specific performance metrics based on (estimated) group-specific learning rates and costs per sample. This flexible approach incorporates preferences of model-builders and other stakeholders, as well as the statistical properties of the learning task. When data collection decisions are made sequentially, we show that under certain conditions this optimization problem can be efficiently solved even without prior knowledge of the learning rates. To illustrate our approach, we conduct a simulation study of polygenic risk scores on synthetic genomic data -- an application domain that often suffers from non-representative data collection. We find that our adaptive sampling strategy outperforms several common data collection heuristics, including equal and proportional sampling, demonstrating the value of strategic dataset design for building equitable models.

We propose a framework for robust evaluation of reasoning capabilities of language models, using functional variants of benchmarks. Models that solve a reasoning test should exhibit no difference in performance over the static version of a problem compared to a snapshot of the functional variant. We have rewritten the relevant fragment of the MATH benchmark into its functional variant MATH(), with functionalization of other benchmarks to follow. When evaluating current state-of-the-art models over snapshots of MATH(), we find a reasoning gap -- the percentage difference between the static and functional accuracies. We find reasoning gaps from 58.35% to 80.31% among the state-of-the-art closed and open weights models that perform well on static benchmarks, with the caveat that the gaps are likely to be smaller with more sophisticated prompting strategies. Here we show that models which anecdotally have good reasoning performance over real-world tasks, have quantifiable lower gaps, motivating the open problem of building "gap 0" models. Code for evaluation and new evaluation datasets, three MATH() snapshots, are publicly available at https://github.com/consequentai/fneval/.

Vision Language Models have recently shown improvements in several objective and verifiable domains such as object detection but continue to underperform on subjective and creative design tasks. A major contributor to this performance gap is the lack of high quality human workflow data that captures a diverse set of preferences and decisions that make human experts good at design tasks. In this work, we first define a unique, expert curated taxonomy of design skills and best practices which we further expand into a set of 126 open ended, subjective, long horizon tasks. Built on top of this and expert solutions, our dataset FigmaTrace contains over 200 hours of human captured video data converted into 3469 design trajectories using a novel design phase-based method. We use our dataset to train four models and show that training on FigmaTrace leads to a performance improvement comparable to frontier closed models such as Claude-Opus-5 and GPT-5.6-Sol on four out of distribution agentic GUI environments. We further perform a useful ablation to attribute these performance improvements to a design phase-based video to trajectory conversion which outperforms prior length-based conversion approaches. Finally, we perform a qualitative analysis on the best performing Qwen3.8-27B outputs to better correlate performance improvements to FigmaTrace's trends. We open source our dataset and the best model for the community.

Cross-encoders distilled from large language models (LLMs) are often more effective re-rankers than cross-encoders fine-tuned on manually labeled data. However, distilled models do not match the effectiveness of their teacher LLMs. We hypothesize that this effectiveness gap is due to the fact that previous work has not applied the best-suited methods for fine-tuning cross-encoders on manually labeled data (e.g., hard-negative sampling, deep sampling, and listwise loss functions). To close this gap, we create a new dataset, Rank-DistiLLM. Cross-encoders trained on Rank-DistiLLM achieve the effectiveness of LLMs while being up to 173 times faster and 24 times more memory efficient. Our code and data is available at https://github.com/webis-de/ECIR-25.

Methodological progress in artificial intelligence (AI) for electronic health records (EHRs) depends on our ability to determine which algorithms work better, and under which conditions. However, such progress is thought to be hindered by difficulties in reproducibility and in defining clinically meaningful evaluation tasks. We empirically study these barriers by re-implementing 12 historical and recent algorithms within a shared evaluation framework and evaluating them on two clinical datasets, MIMIC-IV and NWICU. We compare two complementary task families: expert-authored clinically meaningful tasks and generated tasks defined from randomly sampled event codes and prediction horizons. We ask whether relative algorithms comparisons transfer across task families and datasets, whether residual task heterogeneity contains useful methodological structure, and what a controlled comparison reveals about progress over the last decade. We find that aggregate pairwise comparisons transfer strongly across evaluation settings, including from randomly generated tasks to clinically meaningful tasks and across datasets. At the same time, clinically meaningful tasks exhibit greater task-method interaction, providing preliminary evidence that task properties can help explain when particular modeling choices are advantageous. Finally, newer algorithms do not consistently outperform earlier approaches: gradient-boosted trees remain highly competitive when paired with a modern, wide and sparse representation of the EHR. Together, these results suggest that useful methodological knowledge may require less task engineering than commonly assumed, while highlighting the importance of understanding the structured heterogeneity that remains across tasks and methods.

FinanceQA is a testing suite that evaluates LLMs' performance on complex numerical financial analysis tasks that mirror real-world investment work. Despite recent advances, current LLMs fail to meet the strict accuracy requirements of financial institutions, with models failing approximately 60% of realistic tasks that mimic on-the-job analyses at hedge funds, private equity firms, investment banks, and other financial institutions. The primary challenges include hand-spreading metrics, adhering to standard accounting and corporate valuation conventions, and performing analysis under incomplete information - particularly in multi-step tasks requiring assumption generation. This performance gap highlights the disconnect between existing LLM capabilities and the demands of professional financial analysis that are inadequately tested by current testing architectures. Results show that higher-quality training data is needed to support such tasks, which we experiment with using OpenAI's fine-tuning API. FinanceQA is publicly released at [this https URL](https://huggingface.co/datasets/AfterQuery/FinanceQA).

Preference learning algorithms (e.g., RLHF and DPO) are frequently used to steer LLMs to produce generations that are more preferred by humans, but our understanding of their inner workings is still limited. In this work, we study the conventional wisdom that preference learning trains models to assign higher likelihoods to more preferred outputs than less preferred outputs, measured via ranking accuracy. Surprisingly, we find that most state-of-the-art preference-tuned models achieve a ranking accuracy of less than 60% on common preference datasets. We furthermore derive the idealized ranking accuracy that a preference-tuned LLM would achieve if it optimized the DPO or RLHF objective perfectly. We demonstrate that existing models exhibit a significant alignment gap -- i.e., a gap between the observed and idealized ranking accuracies. We attribute this discrepancy to the DPO objective, which is empirically and theoretically ill-suited to fix even mild ranking errors in the reference model, and derive a simple and efficient formula for quantifying the difficulty of learning a given preference datapoint. Finally, we demonstrate that ranking accuracy strongly correlates with the empirically popular win rate metric when the model is close to the reference model used in the objective, shedding further light on the differences between on-policy (e.g., RLHF) and off-policy (e.g., DPO) preference learning algorithms.

Continual pre-training has increasingly become the predominant approach for adapting Large Language Models (LLMs) to new domains. This process involves updating the pre-trained LLM with a corpus from a new domain, resulting in a shift in the training distribution. To study the behavior of LLMs during this shift, we measured the model's performance throughout the continual pre-training process. we observed a temporary performance drop at the beginning, followed by a recovery phase, a phenomenon known as the "stability gap," previously noted in vision models classifying new classes. To address this issue and enhance LLM performance within a fixed compute budget, we propose three effective strategies: (1) Continually pre-training the LLM on a subset with a proper size for multiple epochs, resulting in faster performance recovery than pre-training the LLM on a large corpus in a single epoch; (2) Pre-training the LLM only on high-quality sub-corpus, which rapidly boosts domain performance; and (3) Using a data mixture similar to the pre-training data to reduce distribution gap. We conduct various experiments on Llama-family models to validate the effectiveness of our strategies in both medical continual pre-training and instruction tuning. For example, our strategies improve the average medical task performance of the OpenLlama-3B model from 36.2% to 40.7% with only 40% of the original training budget and enhance the average general task performance without causing forgetting. Furthermore, we apply our strategies to the Llama-3-8B model. The resulting model, Llama-3-Physician, achieves the best medical performance among current open-source models, and performs comparably to or even better than GPT-4 on several medical benchmarks. We release our models at https://huggingface.co/YiDuo1999/Llama-3-Physician-8B-Instruct.

Commonsense is defined as the knowledge that is shared by everyone. However, certain types of commonsense knowledge are correlated with culture and geographic locations and they are only shared locally. For example, the scenarios of wedding ceremonies vary across regions due to different customs influenced by historical and religious factors. Such regional characteristics, however, are generally omitted in prior work. In this paper, we construct a Geo-Diverse Visual Commonsense Reasoning dataset (GD-VCR) to test vision-and-language models' ability to understand cultural and geo-location-specific commonsense. In particular, we study two state-of-the-art Vision-and-Language models, VisualBERT and ViLBERT trained on VCR, a standard multimodal commonsense benchmark with images primarily from Western regions. We then evaluate how well the trained models can generalize to answering the questions in GD-VCR. We find that the performance of both models for non-Western regions including East Asia, South Asia, and Africa is significantly lower than that for Western region. We analyze the reasons behind the performance disparity and find that the performance gap is larger on QA pairs that: 1) are concerned with culture-related scenarios, e.g., weddings, religious activities, and festivals; 2) require high-level geo-diverse commonsense reasoning rather than low-order perception and recognition. Dataset and code are released at https://github.com/WadeYin9712/GD-VCR.

Repository-level performance-optimization benchmarks such as GSO, SWE-Perf and SWE-fficiency evaluate coding agents by applying patches to real repositories and comparing runtime against unoptimized baselines and official reference patches. Their leaderboard scores are increasingly used as evidence of coding-agent progress, but those scores can conflate runtime instability, benchmark-specific scoring rules, and how many tasks are already solved by at least one public submission. We audit these issues across the three benchmarks. First, we replay the official reference patches for 740 code optimization tasks across four common types of Google Cloud machines. Most benchmark tasks can be replayed, but their reference patches satisfy the original benchmark validity rules in every cross-machine replay for only 39/102 GSO tasks, 11/140 SWE-Perf tasks, and 411/498 SWE-fficiency tasks; SWE-Perf is especially fragile because many reference patches produce close-to-zero runtime changes. Second, we show that public submission rankings depend strongly on the benchmark scoring rule. Among eight public submissions shared by GSO and SWE-fficiency, the official rankings disagree on 9 of 28 pairwise submission comparisons, and SWE-fficiency's leaderboard scoring rule assigns the worst ten tasks overly high score weights of 58.5%-82.8%. Third, looking across 10 public submissions for each task, we find that at least one submission matches or beats the reference patch on 85.3% (384/450) of replay-valid GSO and SWE-fficiency tasks, and beats the unoptimized base code on 99.8% (449/450). Our study complements leaderboard scores by identifying tasks with more reliable performance signals, quantifying per-task score contributions, and exposing the remaining performance gaps that are hidden by aggregate rankings.

Fueled by their remarkable ability to tackle diverse tasks across multiple domains, large language models (LLMs) have grown at an unprecedented rate, with some recent models containing trillions of parameters. This growth is accompanied by substantial computational challenges, particularly regarding the memory and compute resources required for training and fine-tuning. Numerous approaches have been explored to address these issues, such as LoRA. While these methods are effective for fine-tuning, their application to pre-training is significantly more challenging due to the need to learn vast datasets. Motivated by this issue, we aim to address the following questions: Can parameter- or memory-efficient methods enhance pre-training efficiency while achieving performance comparable to full-model training? How can the performance gap be narrowed? To this end, the contributions of this work are the following. (1) We begin by conducting a comprehensive survey that summarizes state-of-the-art methods for efficient pre-training. (2) We perform a benchmark evaluation of several representative memory efficient pre-training approaches to comprehensively evaluate their performance across model sizes. We observe that with a proper choice of optimizer and hyperparameters, full-rank training delivers the best performance, as expected. We also notice that incorporating high-rank updates in low-rank approaches is the key to improving their performance. (3) Finally, we propose two practical techniques, namely weight refactorization and momentum reset, to enhance the performance of efficient pre-training methods. We observe that applying these techniques to the low-rank method (on a 1B model) can achieve a lower perplexity than popular memory efficient algorithms such as GaLore and Fira, while simultaneously using about 25% less memory.

Benchmarks are necessary for healthcare evaluation, but are not sufficient for predicting deployment performance. Our position is that the evaluation--deployment gap arises not because of poorly designed benchmarks, but from implicit assumptions about how users interact with models that cannot be surfaced from benchmarks alone. To make this precise, we propose a classification of assumptions into two categories: task, which can be tested from conversation data alone, and outcome, which requires outcome data and behavioral studies for testing. Critically, outcome assumptions depend on human behavior, something that even well-designed benchmarks cannot directly observe. To demonstrate the operationality of this framework, we retrospectively analyze a healthcare RCT as a case study and find that the gap naturally separates into task and outcome gaps of roughly equal size. To address this, we make two contributions: first, we propose BenchmarkCards, an artifact that documents assumptions, and second, we propose staged evaluation, a procedure that systematically tests assumptions and evaluates performance.

Parameter efficient finetuning has emerged as a viable solution for improving the performance of Large Language Models without requiring massive resources and compute. Prior work on multilingual evaluation has shown that there is a large gap between the performance of LLMs on English and other languages. Further, there is also a large gap between the performance of smaller open-source models and larger LLMs. Finetuning can be an effective way to bridge this gap and make language models more equitable. In this work, we finetune the LLaMA-7B and Mistral-7B models on synthetic multilingual instruction tuning data to determine its effect on model performance on five downstream tasks covering twenty three languages in all. Additionally, we experiment with various parameters, such as rank for low-rank adaptation and values of quantisation to determine their effects on downstream performance and find that higher rank and higher quantisation values benefit low-resource languages. We find that parameter efficient finetuning of smaller open source models sometimes bridges the gap between the performance of these models and the larger ones, however, English performance can take a hit. We also find that finetuning sometimes improves performance on low-resource languages, while degrading performance on high-resource languages.

We study the depth of grade-school math (GSM) problem-solving capabilities of LLMs. To this end, we evaluate their performance on pairs of existing math word problems together so that the answer to the second problem depends on correctly answering the first problem. Our findings reveal a significant reasoning gap in most LLMs, that is performance difference between solving the compositional pairs and solving each question independently. This gap is more pronounced in smaller, more cost-efficient, and math-specialized models. Moreover, instruction-tuning recipes and code generation have varying effects across LLM sizes, while finetuning on GSM can lead to task overfitting. Our analysis indicates that large reasoning gaps are not because of test-set leakage, but due to distraction from additional context and poor second-hop reasoning. Overall, LLMs exhibit systematic differences in their reasoning abilities, despite what their performance on standard benchmarks indicates.

Motivated by the strong performance of CLIP-based models in natural image-text domains, recent efforts have adapted these architectures to medical tasks, particularly in radiology, where large paired datasets of images and reports, such as chest X-rays, are available. While these models have shown encouraging results in terms of accuracy and discriminative performance, their fairness and robustness in the different clinical tasks remain largely underexplored. In this study, we extensively evaluate six widely used CLIP-based models on chest X-ray classification using three publicly available datasets: MIMIC-CXR, NIH-CXR14, and NEATX. We assess the models fairness across six conditions and patient subgroups based on age, sex, and race. Additionally, we assess the robustness to shortcut learning by evaluating performance on pneumothorax cases with and without chest drains. Our results indicate performance gaps between patients of different ages, but more equitable results for the other attributes. Moreover, all models exhibit lower performance on images without chest drains, suggesting reliance on spurious correlations. We further complement the performance analysis with a study of the embeddings generated by the models. While the sensitive attributes could be classified from the embeddings, we do not see such patterns using PCA, showing the limitations of these visualisation techniques when assessing models. Our code is available at https://github.com/TheoSourget/clip_cxr_fairness

Large Language Models are transforming the support for clinical decision and their application in real scenarios. Yet, most benchmarks are conducted in English, and cross-lingual evaluation is needed to tackle the language gaps in global access. We introduce ClinicalBr, the first bilingual benchmark for clinical decision built from real Brazilian case reports. The corpus contains 2,892 cases drawn from 28 SciELO medical journals, spanning 18 specialties, and is structured as parallel Portuguese-English pairs. Each case supports four evaluation tasks: diagnosis retrieval, differential diagnosis, exam recommendation, and treatment planning. We evaluate four models: MedGemma-27B, Sabiá-4, DeepSeek-R1, and o3-mini, across both languages. The central finding is that the Portuguese-English performance gap is task-dependent, not general. In diagnosis retrieval, English yields a consistent advantage across all models, with +7.5-12.1 accuracy points. This advantage disappears in differential diagnosis, exam recommendation, and treatment planning, where confidence intervals cross zero for most models and Portuguese completeness scores are marginally higher. Brazilian-endemic conditions proved easier than the full corpus, not harder, indicating that tropical presentations are adequately represented in current pre-training. Exam recommendation was the hardest task across all models and both languages, with F1 scores below 0.10, well below the differential diagnosis ceiling of 0.20-0.27.

We present NewsQA, a challenging machine comprehension dataset of over 100,000 human-generated question-answer pairs. Crowdworkers supply questions and answers based on a set of over 10,000 news articles from CNN, with answers consisting of spans of text from the corresponding articles. We collect this dataset through a four-stage process designed to solicit exploratory questions that require reasoning. A thorough analysis confirms that NewsQA demands abilities beyond simple word matching and recognizing textual entailment. We measure human performance on the dataset and compare it to several strong neural models. The performance gap between humans and machines (0.198 in F1) indicates that significant progress can be made on NewsQA through future research. The dataset is freely available at https://datasets.maluuba.com/NewsQA.

Strong empirical evidence that one machine-learning algorithm A outperforms another one B ideally calls for multiple trials optimizing the learning pipeline over sources of variation such as data sampling, data augmentation, parameter initialization, and hyperparameters choices. This is prohibitively expensive, and corners are cut to reach conclusions. We model the whole benchmarking process, revealing that variance due to data sampling, parameter initialization and hyperparameter choice impact markedly the results. We analyze the predominant comparison methods used today in the light of this variance. We show a counter-intuitive result that adding more sources of variation to an imperfect estimator approaches better the ideal estimator at a 51 times reduction in compute cost. Building on these results, we study the error rate of detecting improvements, on five different deep-learning tasks/architectures. This study leads us to propose recommendations for performance comparisons.

A paraphrase-quality audit of MathCheck (ICLR 2025) detected 4 semantically incorrect paraphrases in 129 groups (3.1%); removing them drops GPT-4o from rank 2 to rank 4 and elevates Claude Haiku and DeepSeek V3 above it; these ranking changes are invisible to any single-model evaluation. Cross-model unanimity found these errors automatically (>= 3/4 models for MathCheck; >= 6/9 for our primary evaluation) for under $10; in our own dataset the same protocol found that 47% of auto-generated connective-variation paraphrases were semantically incorrect. That flaw compounds a deeper measurement gap: Claude Haiku 4.5 achieves 86% accuracy yet SCR=50%, meaning half its theorems are answered differently under semantically equivalent restatements, while aggregate accuracy across 9 models spans only 86-96% yet Semantic Consistency Rates (SCR) span 50-82% -- a 32-point gap invisible to standard benchmarks. Formally, for any target ranking over 9 frontier models there exists a weighting over paraphrase families that realizes it (No-Free-Benchmark corollary), because no model Pareto-dominates all families -- so benchmark designers who select families are implicitly choosing which model wins. FormInv supplies the audit protocol (replicated on external benchmarks at 100% recall), SCR and per-theorem Cochran's Q as primary invariance measures evaluated on 9 models across 366-811 items (on Lean4-verified theorems), and FormInvSelector for regime-aware model selection.

Comparing human and model performance offers a valuable perspective for understanding the strengths and limitations of embedding models, highlighting where they succeed and where they fail to capture meaning and nuance. However, such comparisons are rarely made, as human performance on embedding tasks is difficult to measure. To fill this gap, we introduce HUME: Human Evaluation Framework for Text Embeddings. While frameworks like MTEB provide broad model evaluation, they lack reliable estimates of human performance, limiting the interpretability of model scores. We measure human performance across 16 MTEB datasets spanning reranking, classification, clustering, and semantic textual similarity across linguistically diverse high- and low-resource languages. Humans achieve an average performance of 77.6% compared to 80.1% for the best embedding model, although variation is substantial: models reach near-ceiling performance on some datasets while struggling on others, suggesting dataset issues and revealing shortcomings in low-resource languages. We provide human performance baselines, insight into task difficulty patterns, and an extensible evaluation framework that enables a more meaningful interpretation of the model and informs the development of both models and benchmarks. Our code, dataset, and leaderboard are publicly available at https://github.com/embeddings-benchmark/mteb.

Similar agent skills can share instructions but differ in their conditions of use. Query-based text selection may retain shared instructions and omit these distinctions. We introduce SkillContrast, a training-free selector that compares retrieved skills and retains their differing text with local context for a pretrained reranker. On 1,235 requests from SameCapRisk-Bench, it yields 54-72 more clean hits (requests that retrieve a helpful skill without its marked risky sibling) than TF-IDF query selection at identical per-candidate input lengths, across 2 retrievers and 2 reranker sizes. Length-matched component replacements identify differing text as the main contributor in the primary setting, with smaller, mixed context effects. Relative to full skill bodies, SkillContrast uses 51.1-58.8% fewer model-input tokens, with 10-18 fewer clean hits at 0.6B and matching or higher observed clean-hit counts at 4B. Candidate-relative differences thus complement query relevance in selecting compact reranking inputs.

Time-dependent data-generating distributions have proven to be difficult for gradient-based training of neural networks, as the greedy updates result in catastrophic forgetting of previously learned knowledge. Despite the progress in the field of continual learning to overcome this forgetting, we show that a set of common state-of-the-art methods still suffers from substantial forgetting upon starting to learn new tasks, except that this forgetting is temporary and followed by a phase of performance recovery. We refer to this intriguing but potentially problematic phenomenon as the stability gap. The stability gap had likely remained under the radar due to standard practice in the field of evaluating continual learning models only after each task. Instead, we establish a framework for continual evaluation that uses per-iteration evaluation and we define a new set of metrics to quantify worst-case performance. Empirically we show that experience replay, constraint-based replay, knowledge-distillation, and parameter regularization methods are all prone to the stability gap; and that the stability gap can be observed in class-, task-, and domain-incremental learning benchmarks. Additionally, a controlled experiment shows that the stability gap increases when tasks are more dissimilar. Finally, by disentangling gradients into plasticity and stability components, we propose a conceptual explanation for the stability gap.

WANDR (Wide ANd Deep Research) is a benchmark of 500 realistic, challenging data-collection tasks for research agents. Each task requires a system to discover a large set of entities that satisfy specified criteria (breadth), investigate each entity through multiple coordinated web searches (depth), and return independently verifiable records with supporting sources and excerpts. Tasks are represented as qualification key hierarchies that specify the entities, relationships, evidence, and required count at each level; a hierarchy with n companies, m employees per company, and k sources per employee requires n x m x k records. This structure supports diverse workflows such as market mapping, due diligence, literature review, product comparison, and talent sourcing, with targets ranging from dozens to thousands of records. WANDR replaces static gold answer sets with task-specific judges that refetch cited pages and verify each record against its evidence, allowing evaluation of current and changing facts. Record verdicts are aggregated into soft and hard precision, recall, and F1 scores that distinguish factual quality, coverage, and hierarchical completeness. The tasks are derived from de-identified product-usage logs and produced through a semi-automated pipeline with automated checks, empirical audits, and human review where needed. We evaluate six production research systems and find that the benchmark is far from saturated: at high effort, the strongest system reaches only 0.363 soft F1 and 0.133 hard F1. Performance degrades as target volume and hierarchy depth increase, with incomplete discovery, missing enrichment, and incomplete evidence construction remaining major bottlenecks. The benchmark and evaluation harness are available at https://github.com/perplexityai/wandr.

Large language models (LLMs) demonstrate exceptional performance on complex reasoning tasks. However, despite their strong reasoning capabilities in high-resource languages (e.g., English and Chinese), a significant performance gap persists in other languages. To investigate this gap in Korean, we introduce HRM8K, a benchmark comprising 8,011 English-Korean parallel bilingual math problems. Through systematic analysis of model behaviors, we identify a key finding: these performance disparities stem primarily from difficulties in comprehending non-English inputs, rather than limitations in reasoning capabilities. Based on these findings, we propose UST (Understand, Solve, and Translate), a method that strategically uses English as an anchor for reasoning and solution generation. By fine-tuning the model on 130k synthetically generated data points, UST achieves a 10.91% improvement on the HRM8K benchmark and reduces the multilingual performance gap from 11.6% to 0.7%. Additionally, we show that improvements from UST generalize effectively to different Korean domains, demonstrating that capabilities acquired from machine-verifiable content can be generalized to other areas. We publicly release the benchmark, training dataset, and models.

Current benchmarks for evaluating Vision Language Models (VLMs) often fall short in thoroughly assessing model abilities to understand and process complex visual and textual content. They typically focus on simple tasks that do not require deep reasoning or the integration of multiple data modalities to solve an original problem. To address this gap, we introduce the PARROT-360V Benchmark, a novel and comprehensive benchmark featuring 2487 challenging visual puzzles designed to test VLMs on complex visual reasoning tasks. We evaluated leading models: GPT-4o, Claude-3.5-Sonnet, and Gemini-1.5-Pro, using PARROT-360V to assess their capabilities in combining visual clues with language skills to solve tasks in a manner akin to human problem-solving. Our findings reveal a notable performance gap: state-of-the-art models scored between 28 to 56 percentage on our benchmark, significantly lower than their performance on popular benchmarks. This underscores the limitations of current VLMs in handling complex, multi-step reasoning tasks and highlights the need for more robust evaluation frameworks to advance the field.

As large language models (LLMs) expand multilingual capabilities, questions remain about the equity of their performance across languages. While many communities stand to benefit from AI systems, the dominance of English in training data risks disadvantaging non-English speakers. To test the hypothesis that such data disparities may affect model performance, this study compares two monolingual BERT models: one trained and tested entirely on Swahili data, and another on comparable English news data. To simulate how multilingual LLMs process non-English queries through internal translation and abstraction, we translated the Swahili news data into English and evaluated it using the English-trained model. This approach tests the hypothesis by evaluating whether translating Swahili inputs for evaluation on an English model yields better or worse performance compared to training and testing a model entirely in Swahili, thus isolating the effect of language consistency versus cross-lingual abstraction. The results prove that, despite high-quality translation, the native Swahili-trained model performed better than the Swahili-to-English translated model, producing nearly four times fewer errors: 0.36% vs. 1.47% respectively. This gap suggests that translation alone does not bridge representational differences between languages and that models trained in one language may struggle to accurately interpret translated inputs due to imperfect internal knowledge representation, suggesting that native-language training remains important for reliable outcomes. In educational and informational contexts, even small performance gaps may compound inequality. Future research should focus on addressing broader dataset development for underrepresented languages and renewed attention to multilingual model evaluation, ensuring the reinforcing effect of global AI deployment on existing digital divides is reduced.

Large language models recall knowledge reliably in English but often fail on the same query posed in a lower-resourced language -- a crosslingual consistency gap that remains underexplored for Indian languages and their code-mixed counterparts. To study this gap, we introduce IndiKLAR, an Indic extension of the KLAR-CLC benchmark covering 18 of the 22 scheduled Indian languages and pairing them with code-mixed variants for 11 widely used language pairs, with native-speaker verification of both monolingual and code-mixed variants for these 11 settings. This three-way alignment offers a unique opportunity to examine how knowledge recall consistency varies across the spectrum of English, code-mixed, and native Indian language inputs. Evaluating across nine open-weight models, we find that the native-language accuracy gap to English can reach sim0.50, while code-mixed inputs close most of it -- bringing performance within sim0.05 of English without any model-level intervention. Motivated by this, we evaluate several prompting strategies that vary in how language conversion is exposed, including a two-stage translate-then-answer setup, a one-stage joint translation-and-answer prompt, and Translate-in-Thought (TinT) -- a single-step strategy in which the model converts the input internally and emits only the final answer. Across the performance trajectory native rightarrow code-mixed rightarrow English, we identify a consistent flip point -- the boundary between incorrect and correct prediction -- that lies between the native and code-mixed settings. Interestingly, this holds whether the trajectory is induced by the input surface form or by the model's internal conversion process.

Training algorithms, broadly construed, are an essential part of every deep learning pipeline. Training algorithm improvements that speed up training across a wide variety of workloads (e.g., better update rules, tuning protocols, learning rate schedules, or data selection schemes) could save time, save computational resources, and lead to better, more accurate, models. Unfortunately, as a community, we are currently unable to reliably identify training algorithm improvements, or even determine the state-of-the-art training algorithm. In this work, using concrete experiments, we argue that real progress in speeding up training requires new benchmarks that resolve three basic challenges faced by empirical comparisons of training algorithms: (1) how to decide when training is complete and precisely measure training time, (2) how to handle the sensitivity of measurements to exact workload details, and (3) how to fairly compare algorithms that require hyperparameter tuning. In order to address these challenges, we introduce a new, competitive, time-to-result benchmark using multiple workloads running on fixed hardware, the AlgoPerf: Training Algorithms benchmark. Our benchmark includes a set of workload variants that make it possible to detect benchmark submissions that are more robust to workload changes than current widely-used methods. Finally, we evaluate baseline submissions constructed using various optimizers that represent current practice, as well as other optimizers that have recently received attention in the literature. These baseline results collectively demonstrate the feasibility of our benchmark, show that non-trivial gaps between methods exist, and set a provisional state-of-the-art for future benchmark submissions to try and surpass.

For more than a decade, researchers have measured progress in object recognition on ImageNet-based generalization benchmarks such as ImageNet-A, -C, and -R. Recent advances in foundation models, trained on orders of magnitude more data, have begun to saturate these standard benchmarks, but remain brittle in practice. This suggests standard benchmarks, which tend to focus on predefined or synthetic changes, may not be sufficient for measuring real world generalization. Consequently, we propose studying generalization across geography as a more realistic measure of progress using two datasets of objects from households across the globe. We conduct an extensive empirical evaluation of progress across nearly 100 vision models up to most recent foundation models. We first identify a progress gap between standard benchmarks and real-world, geographical shifts: progress on ImageNet results in up to 2.5x more progress on standard generalization benchmarks than real-world distribution shifts. Second, we study model generalization across geographies by measuring the disparities in performance across regions, a more fine-grained measure of real world generalization. We observe all models have large geographic disparities, even foundation CLIP models, with differences of 7-20% in accuracy between regions. Counter to modern intuition, we discover progress on standard benchmarks fails to improve geographic disparities and often exacerbates them: geographic disparities between the least performant models and today's best models have more than tripled. Our results suggest scaling alone is insufficient for consistent robustness to real-world distribution shifts. Finally, we highlight in early experiments how simple last layer retraining on more representative, curated data can complement scaling as a promising direction of future work, reducing geographic disparity on both benchmarks by over two-thirds.

Data visualizations are powerful tools for communicating patterns in quantitative data. Yet understanding any data visualization is no small feat -- succeeding requires jointly making sense of visual, numerical, and linguistic inputs arranged in a conventionalized format one has previously learned to parse. Recently developed vision-language models are, in principle, promising candidates for developing computational models of these cognitive operations. However, it is currently unclear to what degree these models emulate human behavior on tasks that involve reasoning about data visualizations. This gap reflects limitations in prior work that has evaluated data visualization understanding in artificial systems using measures that differ from those typically used to assess these abilities in humans. Here we evaluated eight vision-language models on six data visualization literacy assessments designed for humans and compared model responses to those of human participants. We found that these models performed worse than human participants on average, and this performance gap persisted even when using relatively lenient criteria to assess model performance. Moreover, while relative performance across items was somewhat correlated between models and humans, all models produced patterns of errors that were reliably distinct from those produced by human participants. Taken together, these findings suggest significant opportunities for further development of artificial systems that might serve as useful models of how humans reason about data visualizations. All code and data needed to reproduce these results are available at: https://osf.io/e25mu/?view_only=399daff5a14d4b16b09473cf19043f18.

Semantic caching cuts LLM inference costs by serving a cached response to semantically similar queries. Standard practice evaluates these systems using PR-AUC, a metric that only measures how well scores rank and ignores whether they are usable at a fixed threshold. We show this mismatch leads to systematically poor deployment choices, as models with the highest PR-AUC are often the worst in operation. We introduce Precision-Cache Hit Ratio (P-CHR) AUC, a cache-aware metric that measures precision across cache utilization levels, and Calibration Retention Rate (CRR), which captures how much offline ranking quality survives at deployment. We decompose the operational gap between offline and deployed quality into a recoverable calibration component and an irreducible structural component fixed by the dataset's positive rate. Our experiments show that the calibration gap is governed by the training objective rather than data scale, and post-hoc calibration only partially closes it. Ultimately, model selection for semantic caching is a calibration problem, not a ranking one, and measuring it is the first step to closing the gap.

Existing research has identified three structural performance bottlenecks in AI research agents: (1) synchronous single-GPU execution constrains sample throughput, limiting the benefit of search; (2) a generalization gap where validation-based selection causes performance to degrade over extended search horizons; and (3) the limited capability of fixed, single-turn LLM operators imposes a ceiling on search performance. We introduce AIRA_2, which addresses these bottlenecks through three architectural choices: an asynchronous multi-GPU worker pool that increases experiment throughput linearly; a Hidden Consistent Evaluation protocol that delivers a reliable evaluation signal; and ReAct agents that dynamically scope their actions and debug interactively. On MLE-bench-30, AIRA_2 achieves a mean Percentile Rank of 71.8% at 24 hours - surpassing the previous best of 69.9% - and steadily improves to 76.0% at 72 hours. Ablation studies reveal that each component is necessary and that the "overfitting" reported in prior work was driven by evaluation noise rather than true data memorization.

Achieving robust language technologies that can perform well across the world's many languages is a central goal of multilingual NLP. In this work, we take stock of and empirically analyse task performance disparities that exist between multilingual task-oriented dialogue (ToD) systems. We first define new quantitative measures of absolute and relative equivalence in system performance, capturing disparities across languages and within individual languages. Through a series of controlled experiments, we demonstrate that performance disparities depend on a number of factors: the nature of the ToD task at hand, the underlying pretrained language model, the target language, and the amount of ToD annotated data. We empirically prove the existence of the adaptation and intrinsic biases in current ToD systems: e.g., ToD systems trained for Arabic or Turkish using annotated ToD data fully parallel to English ToD data still exhibit diminished ToD task performance. Beyond providing a series of insights into the performance disparities of ToD systems in different languages, our analyses offer practical tips on how to approach ToD data collection and system development for new languages.

Agent skill libraries are becoming routable software assets: a retrieved skill can contribute instructions, scripts, resource bindings, and execution assumptions to an agent. This makes retrieval failures more specific than broad irrelevance. A system can find the right capability family yet expose the wrong same-capability representative. We study this failure as same-capability risk-exposure retrieval. Each benchmark unit pairs a helpful skill with a query-specific risky sibling that shares the capability family but differs on an execution-controlling contract, such as the required resource, precondition, procedure, or artifact. We introduce SameCapRisk-Bench, an auditable benchmark with 1,190 skill-risk units and 1,686 evaluation query cases: 694 marked-sibling units under public library pressure and 496 hard role-flip units where the same two skills swap helpful/risky roles across paired queries. The release records admission evidence, cue/leakage checks, source hashes, family relations, and fixed candidate pools. The benchmark reports helpful ranking together with harmful sibling rate (HSR@K), the top-K exposure of the marked risky sibling. On this benchmark, public SkillRouter, SkillRet, and R3-Skill retrieve helpful skills at high Recall@3 (0.848--0.888) but also expose marked risky siblings frequently (HSR@3 0.346--0.372). A fully public score-and-cluster pipeline lowers HSR@3 to 0.128--0.182, with Recall@3 of 0.713--0.776. Under a benchmark-trained reference scorer, public text-cluster and controlled resolvers reach HSR@3 0.012 and 0.007; the latter attains Recall@3 0.833. Skill retrieval should therefore report both capability matching and same-family risk exposure, with HSR serving as a targeted exposure certificate for fixed skill libraries.

Large language models trained on unrestricted internet corpora inevitably embed information from the future, introducing lookahead bias that compromises the validity of backtests and causal inference in finance and the social sciences. Point-in-time language models--trained exclusively on text available up to each calendar date--eliminate this leakage by construction, but existing efforts typically produce models that lag substantially behind their unconstrained counterparts. We show that this performance gap can be substantially narrowed through scale. Training decoder-only transformers with up to 4 billion parameters on 1 trillion chronologically filtered tokens from FineWeb, we construct a sequence of monthly model checkpoints spanning 2013-2024. Across a range of common-sense reasoning and language understanding benchmarks, our models approach the performance of leading open-weight models of comparable size (e.g., Gemma-3-4B and LLaMA-7B) trained on temporally unrestricted data, although a performance gap remains on several tasks. Instruction fine-tuning via LoRA further improves downstream usability. We release the complete pipeline--including dataset construction, training infrastructure, and evaluation code--to enable reproducible point-in-time language modeling and to support research applications that require strict temporal validity.

A benchmark score compares a system output against a reference, and methodological attention falls almost entirely on the first term. We measure the second. The retained annotation record of a six-language benchmark for personally identifiable information contains two independent annotator labellings, the aggregate shipped as gold, and a reviewer gold from independent expert re-annotation of a sample. Using it, we hold the scored output fixed and exchange the reference. The score moves by 4.95 F1 points for one output and 2.00 for the other (95% CIs [3.23, 6.27] and [0.26, 3.37]), and by 7.55 in the worst language. The comparison between two outputs moves as well: the paired interaction between reference and system is +2.95 points (CI [+1.97, +3.87]), survives correction for multiple testing, and changes one language's margin outright. The cause is an undocumented aggregation default that usually kept one annotator when adjudication did not fire, making the shipped reference a partial copy of an output being scored. We report the resulting reference-sensitivity band, show how to compute one from any retained annotation record, and argue that the quantity belongs beside the score.

Existing benchmarks for frontier models often test specialized, ``PhD-level'' knowledge that is difficult for non-experts to grasp. In contrast, we present a benchmark based on the NPR Sunday Puzzle Challenge that requires only general knowledge. Our benchmark is challenging for both humans and models, however correct solutions are easy to verify, and models' mistakes are easy to spot. Our work reveals capability gaps that are not evident in existing benchmarks: OpenAI o1 significantly outperforms other reasoning models that are on par on benchmarks that test specialized knowledge. Furthermore, our analysis of reasoning outputs uncovers new kinds of failures. DeepSeek R1, for instance, often concedes with ``I give up'' before providing an answer that it knows is wrong. R1 can also be remarkably ``uncertain'' in its output and in rare cases, it does not ``finish thinking,'' which suggests the need for an inference-time technique to ``wrap up'' before the context window limit is reached. We also quantify the effectiveness of reasoning longer with R1 and Gemini Thinking to identify the point beyond which more reasoning is unlikely to improve accuracy on our benchmark.

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 442 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Inference-time scaling can enhance the reasoning capabilities of large language models (LLMs) on complex problems that benefit from step-by-step problem solving. Although lengthening generated scratchpads has proven effective for mathematical tasks, the broader impact of this approach on other tasks remains less clear. In this work, we investigate the benefits and limitations of scaling methods across nine state-of-the-art models and eight challenging tasks, including math and STEM reasoning, calendar planning, NP-hard problems, navigation, and spatial reasoning. We compare conventional models (e.g., GPT-4o) with models fine-tuned for inference-time scaling (e.g., o1) through evaluation protocols that involve repeated model calls, either independently or sequentially with feedback. These evaluations approximate lower and upper performance bounds and potential for future performance improvements for each model, whether through enhanced training or multi-model inference systems. Our extensive empirical analysis reveals that the advantages of inference-time scaling vary across tasks and diminish as problem complexity increases. In addition, simply using more tokens does not necessarily translate to higher accuracy in these challenging regimes. Results from multiple independent runs with conventional models using perfect verifiers show that, for some tasks, these models can achieve performance close to the average performance of today's most advanced reasoning models. However, for other tasks, a significant performance gap remains, even in very high scaling regimes. Encouragingly, all models demonstrate significant gains when inference is further scaled with perfect verifiers or strong feedback, suggesting ample potential for future improvements.

We introduce LegalBench-BR, the first public benchmark for evaluating language models on Brazilian legal text classification. The dataset comprises 3,105 appellate proceedings from the Santa Catarina State Court (TJSC), collected via the DataJud API (CNJ) and annotated across five legal areas through LLM-assisted labeling with heuristic validation. On a class-balanced test set, BERTimbau-LoRA, updating only 0.3% of model parameters, achieves 87.6% accuracy and 0.87 macro-F1 (+22pp over Claude 3.5 Haiku, +28pp over GPT-4o mini). The gap is most striking on administrativo (administrative law): GPT-4o mini scores F1 = 0.00 and Claude 3.5 Haiku scores F1 = 0.08 on this class, while the fine-tuned model reaches F1 = 0.91. Both commercial LLMs exhibit a systematic bias toward civel (civil law), absorbing ambiguous classes rather than discriminating them, a failure mode that domain-adapted fine-tuning eliminates. These results demonstrate that general-purpose LLMs cannot substitute for domain-adapted models in Brazilian legal classification, even when the task is a simple 5-class problem, and that LoRA fine-tuning on a consumer GPU closes the gap at zero marginal inference cost. We release the full dataset, model, and pipeline to enable reproducible research in Portuguese legal NLP.

We investigate how teacher-student capacity relationships modulate knowledge distillation (KD) effectiveness in ResNet-based image classification on CIFAR-10. Across three teacher-student pairs -- R50->R18, R34->R18, and R50->R34 -- we compare Logit-KD and Feature-KD under controlled, reproducible conditions (3 seeds, mean+/-std reported throughout). We report three main findings. First, student capacity is a key moderating factor in distillation gain: R34 students benefit substantially more from KD than R18 students even when teacher-student accuracy gaps are comparable, with the strongest gain of +0.30pp observed for R50->R34 Feature-KD versus +0.18pp for R34->R18 Feature-KD and +0.00pp for R34->R18 Logit-KD. Second, implementation correctness critically affects Feature-KD: a gradient clipping bug that excluded projection layers suppressed Feature-KD performance and produced misleading comparisons with Logit-KD. After correction, Feature-KD matches or outperforms Logit-KD in two of three pairs, reaching 95.55% on R50->R34 against a baseline of 95.25%. Third, input-resolution-aware architecture is a prerequisite for effective distillation: correcting the ResNet stem for 32x32 inputs raises teacher accuracy by over 5pp -- an order of magnitude larger than any KD gain. All code and results are available at github.com/umutonuryasar/kd-capacity-gap.

Learning from experience is essential for LLM agents to adapt to unfamiliar and dynmaic environments. Evaluating this ability is therefore important for understanding how effectively agents acquire and use new knowledge. Existing benchmarks have sought to evaluate this ability, but they primarily evaluate tasks whose rules are provided in the instructions or already familiar to pretrained models, making it difficult to distinguish learning from interactions from reasoning with existing knowledge. To address this, we introduce Learn2Play Bench, a benchmark of newly designed text-based games, whose rules are novel or counterintuitive, requiring agents to acquire knowledge through interaction rather than rely solely on pretrained knowledge. These games provide reproducible feedback and automatic scoring, enabling controlled evaluation of learning across repeated attempts. We also vary game instances to test whether agents can apply what they have learned to new situations. Therefore, we evaluate how backbone models, self-evolving methods, and agent harnesses affect agents' learning ability, revealing three findings: (1) Experience retention: Retaining complete records of actions and feedback can support more effective learning than summarizing these experiences into rules or strategies. (2) Human agent gap: Top-performing human players achieve higher peak scores than the evaluated agents. Human explore more varied strategies, and repeat actions less. (3) Harness matters: With the backbone fixed, changing the harness can improve performance while reducing estimated inference cost. Together, these findings provide insights into how LLM agents learn from experience and suggest directions for future work to improve their learning ability. Project website: https://liushiliushi.github.io/learn2play-bench-website/

A small language model can receive the governing statutory provision and still answer incorrectly. We test whether fine-tuning on examples containing relevant law improves later use of retrieved law. We curate 2{,}165 bilingual QA records from six Bangladeshi acts and three schedules, then fine-tune Qwen3.5 at 0.8B, 2B, and 4B. Evaluation uses the 2022 and 2023 Bangladesh Bar Council exams in Bangla and machine-translated English, with no retrieval, BM25, or FAISS, scored by strict consistency over three seeded runs. At 0.8B, fine-tuning raises the 2022 English FAISS score from 2 to 34 of 100. Gains at 0.8B and 2B survive paired testing, but the 4B model has no detectable net gain: Bangla improves while several English conditions regress. Fine-tuning also reduces answers that drift from Bangla into mostly English from 44.0--53.2\% to 0.2--0.7\%, with adjusted p<.001 at every scale. Retrieval quality is therefore not the only bottleneck. Small bilingual legal models also differ in how they use supplied law and whether they answer in the requested language. The dataset is publicly available at https://huggingface.co/datasets/momahadi/bangladesh-legal-qa-dataset.

Safety benchmarks assume that test-condition behavior predicts deployment behavior, an assumption that fails if models detect evaluation cues and adapt. This opens a gap between benchmark performance and deployment behavior: compliance measured under test conditions becomes an optimistic upper bound that overstates how safely a model behaves once the evaluation harness is removed. We characterize this evaluation awareness through eight experiments across 37 open-weight models and seven families. (i)Detection is moderate and training-driven (24/37 models exceed chance, best AUROC 0.714 vs.0.819 human, with instruction tuning dominating over scale). (ii)Detection shifts safety behavior (hard refusal drops 5.8 percentage points under hypothetical framing, and 21/140 HarmBench framing effects are significant, with compliance rising up to +30 percentage points. (iii)Representations survive behavioral collapse (probes retain AUROC 0.98 under rewrites that drive behavior below chance, and multi-layer steering causally moves three downstream tasks while random controls do not). (iv)These axes are weakly coupled (only 1/15 correlations are significant, the sole robust link being behavioral detection versus framing resistance, ρ=-0.79, p<0.001). We call this gap the benchmark illusion: because detectability, behavioral manifestation, and controllability vary independently, it is multivariate rather than a single number, so no single awareness score is a reliable proxy for deployment safety.

Being able to provide explanations for a model's decision has become a central requirement for the development, deployment, and adoption of machine learning models. However, we are yet to understand what explanation methods can and cannot do. How do upstream factors such as data, model prediction, hyperparameters, and random initialization influence downstream explanations? While previous work raised concerns that explanations (E) may have little relationship with the prediction (Y), there is a lack of conclusive study to quantify this relationship. Our work borrows tools from causal inference to systematically assay this relationship. More specifically, we study the relationship between E and Y by measuring the treatment effect when intervening on their causal ancestors, i.e., on hyperparameters and inputs used to generate saliency-based Es or Ys. Our results suggest that the relationships between E and Y is far from ideal. In fact, the gap between 'ideal' case only increase in higher-performing models -- models that are likely to be deployed. Our work is a promising first step towards providing a quantitative measure of the relationship between E and Y, which could also inform the future development of methods for E with a quantitative metric.

Human supervisors rely on coding agents' reports to assess progress, verify repairs, and decide when to intervene. But a supervisor cannot investigate a failure they were never told occurred. Failed attempts, workarounds, and unresolved risks may disappear from routine reports, leaving supervisors unaware of issues that warrant scrutiny. We conduct CTXGAP, a measurement study quantifying the gap in task-state recoverability between agents' accumulated user-facing reports and their full recorded execution contexts. Across 37 coding tasks and 4,799 valid probes, agents answer 93.6% of task-state questions on demand; full-context readers recover 92.7%, while report-only readers recover 52.0%, a 40.7-point gap. At the final checkpoints of 30 verified fixes, the gap remains 34.8 points. Human annotation identifies states mentioned without enough supporting detail as the largest category of unrecovered information, and model-assessed supervision-relevant states retain substantial gaps. Full terminal scrollback improves recovery to 73.6% but still trails full context. Under matched report lengths, context-informed rewrites improve recovery by about eight points, without a detectable targeted-over-natural advantage. Positive gaps persist across harnesses and independent DeepSeek execution and measurement. CTXGAP establishes report-based task-state recoverability as a complementary evaluation dimension.

Standard evaluation protocols reveal a counterintuitive phenomenon: on 7.7% of benchmark problems spanning five datasets, larger language models underperform smaller ones by 28.4 percentage points despite 10-100x more parameters. Through systematic evaluation of 31 models (0.5B-405B parameters) across 1,485 problems, we identify the mechanism as spontaneous scale-dependent verbosity that introduces errors through overelaboration. Causal intervention experiments demonstrate this reflects correctable prompt design rather than fundamental capability limitations. Constraining large models to produce brief responses improves accuracy by 26 percentage points and reduces performance gaps by up to two-thirds. Most critically, brevity constraints completely reverse performance hierarchies on mathematical reasoning and scientific knowledge benchmarks, with large models achieving 7.7-15.9 percentage point advantages over small models -- direct inversions of the original gaps. These reversals prove large models possess superior latent capabilities that universal prompting masks. We validate findings through three independent contamination tests and demonstrate inverse scaling operates continuously across the full parameter spectrum, with dataset-specific optimal scales ranging from 0.5B to 3.0B parameters. Our results establish that maximizing large model performance requires scale-aware prompt engineering rather than universal evaluation protocols, with immediate implications for deployment: prompt adaptation simultaneously improves accuracy and reduces computational costs.

The Abstraction and Reasoning Corpus (ARC-AGI) has become a key benchmark for fluid intelligence in AI. This survey presents the first cross-generation analysis of 82 approaches across three benchmark versions and the ARC Prize 2024-2025 competitions. Our central finding is that performance degradation across versions is consistent across all paradigms: program synthesis, neuro-symbolic, and neural approaches all exhibit 2-3x drops from ARC-AGI-1 to ARC-AGI-2, indicating fundamental limitations in compositional generalization. While systems now reach 93.0% on ARC-AGI-1 (Opus 4.6), performance falls to 68.8% on ARC-AGI-2 and 13% on ARC-AGI-3, as humans maintain near-perfect accuracy across all versions. Cost fell 390x in one year (o3's 4,500/task to GPT-5.2's 12/task), although this largely reflects reduced test-time parallelism. Trillion-scale models vary widely in score and cost, while Kaggle-constrained entries (660M-8B) achieve competitive results, aligning with Chollet's thesis that intelligence is skill-acquisition efficiency. Test-time adaptation and refinement loops emerge as critical success factors, while compositional reasoning and interactive learning remain unsolved. ARC Prize 2025 winners needed hundreds of thousands of synthetic examples to reach 24% on ARC-AGI-2, confirming that reasoning remains knowledge-bound. This first release of the ARC-AGI Living Survey captures the field as of February 2026, with updates at https://nimi-ai.com/arc-survey/

Code benchmarks such as HumanEval are widely adopted to evaluate the capabilities of Large Language Models (LLMs), providing insights into their strengths and weaknesses. However, current benchmarks primarily exercise LLMs' capability on common coding tasks (e.g., bubble sort, greatest common divisor), leaving domain-specific coding tasks (e.g., computation, system, cryptography) unexplored. To fill this gap, we propose a multi-domain code benchmark, DOMAINEVAL, designed to evaluate LLMs' coding capabilities thoroughly. Our pipeline works in a fully automated manner, enabling a push-bottom construction from code repositories into formatted subjects under study. Interesting findings are observed by evaluating 12 representative LLMs against DOMAINEVAL. We notice that LLMs are generally good at computation tasks while falling short on cryptography and system coding tasks. The performance gap can be as much as 68.94% (80.94% - 12.0%) in some LLMs. We also observe that generating more samples can increase the overall performance of LLMs, while the domain bias may even increase. The contributions of this study include a code generation benchmark dataset DOMAINEVAL, encompassing six popular domains, a fully automated pipeline for constructing code benchmarks, and an identification of the limitations of LLMs in code generation tasks based on their performance on DOMAINEVAL, providing directions for future research improvements. The leaderboard is available at https://domaineval.github.io/.

Data-centric curation pipelines frequently rely on model confidence scores to flag and filter noisy or mislabeled training instances. Evaluating this filtering convention on a large-scale consumer lending sample (LendingClub, N = 1,344,936) uncovers an underlying demographic asymmetry: high-income defaulters are disproportionately classified as label noise relative to low-income defaulters (Cramer's V approximately 0.03-0.07). Re-examining this behavior through the lens of equal opportunity [Hardt et al., 2016] reveals a far more severe discrepancy: a 16.86 percentage point gap in true positive rate (recall) between high- and low-income borrowers who ultimately defaulted. Implementing a sequential feature-blinding methodology allows us to isolate the drivers of this disparity across three distinct mechanisms: (1) direct reliance on self-reported applicant income; (2) algorithmic absorption of upstream institutional bias encoded within origination interest rates; and (3) a residual disparity (3.55 percentage points in cross-validation; 2.56 percentage points on a held-out test partition, Z = -4.04, p < 0.0001) that remains even after purging both income and interest rates from the model. Out-of-sample signed SHAP valuations demonstrate that this residual gap is maintained by structural proxies, most notably loan amount and home ownership status. These empirical findings show that simply blinding an algorithm to sensitive attributes fails to ensure fairness when institutional pricing decisions and behavioral proxy variables collectively reconstruct the omitted signals. We outline the practical implications of these findings for auditing data-centric AI workflows within regulated financial institutions.

Streaming video understanding requires models to interpret evidence as it arrives, yet current evaluations often report task scores without specifying when evidence becomes valid, how visual history is maintained, or how responses are triggered. As a result, similar scores may correspond to different workloads, failure modes, and operational behavior. We introduce TRACE (Temporal Audit and Condition-aware Evaluation), a condition-aware benchmark and evaluation framework that makes these factors explicit. TRACE combines temporally audited visual tasks with evidence timing and instruction-dependent trigger annotations, a unified causal Core--Adapter protocol that controls information availability while recording actual history processing and response events, and multidimensional reporting of answer quality, timeliness, response-selection behavior, workload, completion, and reliability. On 1,240 records from 517 videos, we evaluate eight publicly available models or systems in eight configurations. We find that nearly identical QA accuracy can mask substantial differences in completion, answer validity, and generation workload, while proactive performance separates into response quality, response delay, false alarms (responses emitted while no target window is currently valid and a later one remains), and missed target windows. These results show that streaming-video performance should be interpreted as execution-conditioned system behavior rather than a single score. Our benchmark and code can be accessed at https://github.com/om-ai-lab/trace-bench{https://github.com/om-ai-lab/trace-bench}.

While current Computer Use Agent (CUA) benchmarks measure task completion effectively, they provide limited assessment of enterprise deployment readiness, emphasizing functional correctness over the operational reliability required for production systems. We present UI-CUBE (UiPath Computer Use BEnchmark), a systematic benchmark comprising 226 tasks across two difficulty tiers designed to expose fundamental architectural limitations in current CUAs. Our evaluation covers simple UI interactions (136 tasks) and complex workflows including copy-paste tasks (50 tasks) and enterprise application scenarios (40 tasks), with systematic interface variation coverage, multi-resolution testing and automated validation of task success through the application state. Evaluation of five state-of-the-art models reveals a sharp capability cliff rather than gradual performance degradation. Simple UI interactions achieve 67-85% success rates (compared to 97.9% human performance), but complex workflows drop precipitously to 9-19%. Human evaluators with no prior application experience achieve only 61.2% on complex tasks despite near-perfect performance on simple tasks, establishing realistic performance ceilings. This discontinuous performance pattern -- where agents achieve 68-87% of human performance on simple tasks but only 15-32% on complex workflows -- indicates fundamental architectural limitations in memory management, hierarchical planning, and state coordination rather than incremental capability gaps addressable through better training or prompting. UI-CUBE functions as an enterprise-readiness diagnostic, revealing that while current CUAs can manipulate individual interface elements, they cannot yet function as reliable workflow automation tools. These findings provide architectural insights essential for developing production-ready CUAs capable of managing complex, multi-step enterprise processes.

Large language models (LLMs) have the potential to boost human productivity by speeding up task completion -- provided users know when to offload cognitive work to them. But we do not know if users are well-calibrated in estimating these potential time savings. We conducted a preregistered large-scale behavioral study (N = 1237) to characterize mismatches between expectations and reality, with a focus on simple cognitive tasks. While actual completion times between independent completion and AI-assisted completion did not differ, participants predicted AI to be significantly faster. The same bias was not observed when imagining help from another human participant. We identify a speedup illusion where people have accurate forecasts of independent completion times but significantly underestimate AI-assisted times. Additionally, time and effort dissociate: participants reported lower subjective effort with AI despite equivalent completion times. This suggests that completion time itself is not sufficient to characterize efficiency gains.

State-of-the-art AI systems can be significantly improved without expensive retraining via "post-training enhancements"-techniques applied after initial training like fine-tuning the system to use a web browser. We review recent post-training enhancements, categorizing them into five types: tool-use, prompting methods, scaffolding, solution selection, and data generation. Different enhancements improve performance on different tasks, making it hard to compare their significance. So we translate improvements from different enhancements into a common currency, the compute-equivalent gain: how much additional training compute would be needed to improve performance by the same amount as the enhancement. Our non-experimental work shows that post-training enhancements have significant benefits: most surveyed enhancements improve benchmark performance by more than a 5x increase in training compute, some by more than 20x. Post-training enhancements are relatively cheap to develop: fine-tuning costs are typically <1% of the original training cost. Governing the development of capable post-training enhancements may be challenging because frontier models could be enhanced by a wide range of actors.

Calibration is a fundamental property of a good predictive model: it requires that the model predicts correctly in proportion to its confidence. Modern neural networks, however, provide no strong guarantees on their calibration -- and can be either poorly calibrated or well-calibrated depending on the setting. It is currently unclear which factors contribute to good calibration (architecture, data augmentation, overparameterization, etc), though various claims exist in the literature. We propose a systematic way to study the calibration error: by decomposing it into (1) calibration error on the train set, and (2) the calibration generalization gap. This mirrors the fundamental decomposition of generalization. We then investigate each of these terms, and give empirical evidence that (1) DNNs are typically always calibrated on their train set, and (2) the calibration generalization gap is upper-bounded by the standard generalization gap. Taken together, this implies that models with small generalization gap (|Test Error - Train Error|) are well-calibrated. This perspective unifies many results in the literature, and suggests that interventions which reduce the generalization gap (such as adding data, using heavy augmentation, or smaller model size) also improve calibration. We thus hope our initial study lays the groundwork for a more systematic and comprehensive understanding of the relation between calibration, generalization, and optimization.

We investigate large language model performance across five orders of magnitude of compute scaling in eleven recent model architectures. We show that average benchmark performance, aggregating over many individual tasks and evaluations as in the commonly-used BIG-Bench dataset, is decently predictable as a function of training compute scale. Specifically, when extrapolating BIG-Bench Hard performance across one order of magnitude in compute, we observe average absolute errors of 6 percentage points (pp). By contrast, extrapolation for individual BIG-Bench tasks across an order of magnitude in compute yields higher average errors of 18pp. Nonetheless, individual task performance remains significantly more predictable than chance. Overall, our work suggests compute scaling provides a promising basis to forecast AI capabilities in diverse benchmarks, though predicting performance in specific tasks poses challenges.

In cognitive science, resource rationality asks how an agent should allocate limited computation to maximize expected value. Most reasoning and agent benchmarks use independent per-task budgets; existing shared-budget studies do not calibrate suite performance against the same model's demonstrated single-problem competence. We introduce R^3-Bench, which evaluates six-problem suites under shared budgets across mathematics, competitive programming, and abstract reasoning in tool-free and agentic settings. Matched single-problem response curves define an offline empirical oracle over observed successes. Across 72 main-table cells for six models, the oracle mean matches or exceeds the contest mean in all cells and is strictly higher in 71. Under moderate tool-free pressure, equal-allocation replay also exceeds contest performance for four of six models. Trajectory diagnostics reveal limited strategy updating and pressure-dependent failure patterns. In a three-model diagnostic under strong agentic pressure, at least one fixed scheduler exceeds the contest mean in six of nine cells, but no policy dominates across domains. These results expose a persistent gap between demonstrated competence and shared-budget realization.

Instruction-tuning data are judged by quality metrics, and tuned models are judged by benchmarks, but both judgments pass through an output interface: the surface format in which an answer is written. Using gradient signatures across 12 tasks, four semantically equivalent interfaces, three model families, and controlled corruptions, we show that this interface confounds both measurements. Spectral statistics such as effective rank are provably invariant to interface rotation and empirically blind to semantic corruption, while the direction of the update carries the quality signal. The interface-varying residual is not noise: it identifies each unit's own target task perfectly across all three families. Capability itself is stored relative to the training interface: a skill that raises accuracy by more than 40 points under the training format can be nearly invisible under every other, and correcting a single generation budget flips the measured effect of fine-tuning on GSM8K from a gain into a large loss. Pre-registered interventions delimit where this geometry stops short of control. Data quality and model capability are interface-conditioned quantities, and current practice often reports the interface instead of the content.

The scientific scale-up of large language models (LLMs) necessitates a comprehensive understanding of their scaling properties. However, the existing literature on the scaling properties only yields an incomplete answer: optimization loss decreases predictably as the model size increases, in line with established scaling law; yet no scaling law for task has been established and the task performances are far from predictable during scaling. Task performances typically show minor gains on small models until they improve dramatically once models exceed a size threshold, exemplifying the ``emergent abilities''. In this study, we discover that small models, although they exhibit minor performance, demonstrate critical and consistent task performance improvements that are not captured by conventional evaluation strategies due to insufficient measurement resolution. To measure such improvements, we introduce PassUntil, an evaluation strategy through massive sampling in the decoding phase. We conduct quantitative investigations into the scaling law of task performance. Firstly, a strict task scaling law is identified, enhancing the predictability of task performances. Remarkably, we are able to predict the performance of the 2.4B model on code generation with merely 0.05\% deviation before training starts. Secondly, underpinned by PassUntil, we observe concrete evidence of emergent abilities and ascertain that they are not in conflict with the continuity of performance improvement. Their semblance to break-through is that their scaling curve cannot be fitted by standard scaling law function. We then introduce a mathematical definition for the emergent abilities. Through the definition, we refute a prevalent ``multi-step reasoning hypothesis'' regarding the genesis of emergent abilities and propose a new hypothesis with a satisfying fit to the observed scaling curve.

Changes produced by coding agents can pass functional tests while leaving repository contribution requirements unmet. Following repository-specific norms requires identifying guidance dispersed across repository sources and interpreting its conditions and exceptions. Retrieval and documentation approaches supply general context, but agents must still determine which norms apply. We introduce RepoNorm to acquire explicit and implicit repository norms independently of coding tasks. It checks norm content and applicability using repository evidence, consults Git history when needed, and delivers norm packages to existing coding agents. Our evaluation uses three coding models and 121 tasks from RepoNormBench. Against the baseline with no additional generated guidance (Raw), relative improvements are 7.42-10.77% for Overall Norm Compliance Rate (NCR), 31.64-45.44% for Contribution NCR, and 11.34-17.69% for Prompt-omitted NCR. All three coding models also obtain higher values on these NCR measures with RepoNorm than with CodeWiki documentation. Functional success rates show observed gains of 5.79-9.09 percentage points over Raw; the paired comparisons do not reach the significance threshold. Sampled precision is 86% with RepoNorm's default configuration.

AI agents increasingly operate in environments where they can diagnose failures and improve through experience, yet existing evaluations largely measure what an agent can do at a fixed point in time rather than how effectively it learns. Evaluating self-improvement requires answering three questions: does future performance improve and generalize beyond the interactions that enabled learning; how efficiently are new capabilities acquired; and where does the self-improvement process break down? To answer these questions, we study self-improvement in a controlled setting where agents amortize past experience into reusable artifacts that are inherited by future instances. At each checkpoint, we measure performance on training and held-out environment interactions while accounting for learning cost. We introduce agent plasticity, the efficiency with which an agent converts experience into gains in future held-out performance. Across multiple environments, frontier models exhibit sharply different improvement trajectories despite comparable opportunities to learn. Some achieve substantial and persistent gains, while others remain near or below their initial performance, and gains within the training regime often transfer only partially to out-of-distribution conditions. Endpoint capability and acquisition efficiency also diverge: the agent that ultimately performs best need not be the one that improves most efficiently. Tracing failures through the improvement loop further reveals different candidate bottlenecks. Agents with low plasticity often fail to reuse relevant artifacts, whereas more plastic agents may still fail despite reusing relevant artifacts, pointing to limitations in artifact quality, generalization, or application. Evaluating self-improving agents requires measuring not only what they can do, but how effectively they become better through experience.

A benchmark score means nothing without knowing what a trivial method achieves and what the best possible method could achieve. We construct both bounds for a task with a rare kind of ground truth: predicting which sentences a crowd of readers -- highlighting for their own purposes, unpaid, uninstructed, and blind to each other -- marked in 120 web documents. The floor is naive truncation (lead); the ceiling is a split-half oracle: half the crowd predicting the other half. The gap between them is +0.2028 AP [+0.1698, +0.2342, domain-clustered], and three findings structure it. First, the gap is semantic: position and length features recover 5% of it. Second, frontier language models reach 35-53% of it zero-shot -- far above classical baselines, far below the crowd; a state-of-the-art prompt compressor (LLMLingua-2) lands below the floor, indistinguishable from random selection. Third, an unweighted cross-vendor fusion of five frontier rankings plus a position prior reaches 60%, beating the best single model by +0.0159 [+0.0044, +0.0269; Holm p=0.019] -- a gain that survives ablation of its best member, split-half arm selection, prompt paraphrase, and label, gate, and seed perturbations, and was CONFIRMED by a pre-registered replication on 217 independent documents (+0.0179, Holm p=0.042). Finally, the bracket compresses: distilling the fusion into one open-weight 8B student that reads the whole document retains 90% of the fusion's edge and reaches statistical parity with the strongest single frontier model (+0.0070 [-0.0068, +0.0200]), where a local-context student retains only 63% -- the crowd's signal lives in document-level structure, and the cheapest known improvement is to ask several different models and average.

Deploying large language models for legal question answering raises challenges that general-purpose leaderboards do not capture, particularly for low-resource languages and under hard operational constraints. We report on building and operating a retrieval-augmented (RAG) legal assistant for Uzbek that must run in two regimes: a managed cloud service that maximizes answer quality within a per-token cost ceiling, and an on-premises deployment for clients whose legal data may not leave their infrastructure, restricting us to open-weight models on limited local hardware under latency constraints. Because no evaluation existed for this setting, we build two domain benchmarks: a retrieval benchmark of 178 expert-annotated legal queries with gold provision spans, and an end-to-end benchmark of 504 expert-curated question--answer pairs scored by an LLM judge whose ratings we validate against human judgments and against an independent-family judge. Applying these benchmarks under each regime, we find the open-versus-proprietary gap is small and cheaply closed by fine-tuning. Therefore, we train UTE-1, which is a state-of-the-art text embedder among open models for Uzbek. We also demonstrate that closing the performance gap via fine-tuning is both impractical due to the intensive hardware demands of long-context legal Q\&A and unnecessary, given that legal acts change frequently. We support this by reporting a negative result from a QLoRA experiment. We distill practical guidance for similar deployments, drawn from a system serving real users in production. We release our benchmarks, evaluation code and the fine-tuned embedder (UTE-1) https://metric-ai-lab.github.io/Uzbek-Legal-RAG/{at this https URL} to support future work on low-resource legal NLP.

Evaluation of large language models (LLMs) for code has primarily relied on static benchmarks, including HumanEval (Chen et al., 2021), which measure the ability of LLMs to generate complete code that passes unit tests. As LLMs are increasingly used as programmer assistants, we study whether gains on existing benchmarks translate to gains in programmer productivity when coding with LLMs, including time spent coding. In addition to static benchmarks, we investigate the utility of preference metrics that might be used as proxies to measure LLM helpfulness, such as code acceptance or copy rates. To do so, we introduce RealHumanEval, a web interface to measure the ability of LLMs to assist programmers, through either autocomplete or chat support. We conducted a user study (N=213) using RealHumanEval in which users interacted with six LLMs of varying base model performance. Despite static benchmarks not incorporating humans-in-the-loop, we find that improvements in benchmark performance lead to increased programmer productivity; however gaps in benchmark versus human performance are not proportional -- a trend that holds across both forms of LLM support. In contrast, we find that programmer preferences do not correlate with their actual performance, motivating the need for better, human-centric proxy signals. We also open-source RealHumanEval to enable human-centric evaluation of new models and the study data to facilitate efforts to improve code models.

Existing agent benchmarks typically evaluate agents on well-specified tasks defined by issue reports, specifications, or task tickets. Real-world customer delivery is different: forward-deployed engineers (FDEs) must clarify ambiguous requirements, work in live systems, and uncover unstated constraints. We introduce FDE-Gym, a benchmark of 95 end-to-end delivery tasks from 20 customer scenarios across 18 sectors, each adapted by industry experts from real-world delivery projects. In simulated customer environments with scripted stakeholders and modifiable customer systems, agents interact with stakeholders and operate on production-like systems to deliver working solutions. FDE-Gym scores a delivery by how much it improves the customer's own success metric, normalized between 0 (unchanged system) and 1 (reference solution), with penalties for stakeholder interactions (messages, trial runs, and waiting time). An online platform lets humans practice FDE skills under the same grading criteria. We evaluate 13 model configurations from seven providers. In the realistic setting, where unstated constraints must be discovered, the best configuration achieves a mean FDE-Gym score of 0.56 and resolves only 14 of the 95 tasks (a score of at least 0.8 with no incident). Discovering unstated constraints is the primary bottleneck: performance drops sharply when models must uncover them without explicit guidance. Yet information discovery alone does not account for the gap: even on the 23 tasks where the best configuration obtains every key fact from stakeholders, it reaches a normalized business uplift of 0.8 or above on only 10. These results reveal a substantial gap between solving well-specified tasks and delivering effective AI systems for real customers.

Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the Open LLM Leaderboard aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly). However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed. We use data from n > 5000 LLMs to identify the most informative items of six benchmarks, ARC, GSM8K, HellaSwag, MMLU, TruthfulQA and WinoGrande (with d=28,632 items in total). From them we distill a sparse benchmark, metabench, that has less than 3% of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities. We show that these estimators (1) can be used to reconstruct each original individual benchmark score with, on average, 1.5% root mean square error (RMSE), (2) reconstruct the original total score with 0.8% RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is r = 0.93.

While question-answering~(QA) benchmark performance is an automatic and scalable method to compare LLMs, it is an indirect method of evaluating their underlying problem-solving capabilities. Therefore, we propose a holistic and generalizable framework based on cascaded question disclosure that provides a more accurate estimate of the models' problem-solving capabilities while maintaining the scalability and automation. This approach collects model responses in a stagewise manner with each stage revealing partial information about the question designed to elicit generalized reasoning in LLMs. We find that our approach not only provides a better comparison between LLMs, but also induces better intermediate traces in models compared to the standard QA paradigm. We empirically verify this behavior on diverse reasoning and knowledge-heavy QA datasets by comparing LLMs of varying sizes and families. Our approach narrows the performance gap observed in the standard QA evaluation settings, indicating that the prevalent indirect QA paradigm of evaluation overestimates the differences in performance between models. We further validate our findings by extensive ablation studies.

Whether a language model looks demographically biased can depend on how the audit asks its question. A charitable-aid benchmark reports that the same models favor minority applicants when rating requests one at a time and penalize some when ranking side by side. We test whether that reversal generalizes to hiring, lending, and medical triage: 40,726 requests to five models, applications differing only in the applicant's name, and a primary test fixed before collection. It does not. None of 36 planned contrasts survives correction. The rating advantage keeps its sign at roughly half the published size, and a precision extension bounds any hiring ranking penalty below the published effect, though the lending and triage ranking floors sit above that margin, so the exclusion is conclusive for hiring ranking and for rating in all three domains only. Planted disparities tracking their injected sizes and a directional replication on the original aid materials bound these nulls. The audit is livelier than the demographics: models recognize transparent audits nearly always, tie every identical-content comparison whether the varying detail is race or a hobby, and reward first-listed candidates as much as any demographic effect we measure. Audit verdicts reflect audit construction more than demographic bias.

Probes are the workhorse of interpretability. If a model's hidden states predict a variable, the model is said to represent it. But a probe score has no fixed meaning. An R^2 of 0.6 may only reflect what the input already gives away, and the same score can mean different things on different data. We propose reading every probe score against two reference points: a floor, what a declared set of simple inputs already predicts, and a ceiling, what the full input can predict. The gap between them, the headroom, is the range in which a probe can show that a model computes something beyond the simple inputs. We prove that headroom vanishes in two ways: the target stops depending on a hidden variable the model must infer, or the input stops revealing it. We test this on transformers trained for in-context meta-analysis, which must infer the hidden heterogeneity between studies to weight them correctly, and where both reference points are known. Under distribution shift, probe scores fall and prediction error rises 12--15times, yet the model recovers a similar share of the headroom, indicating that the data lost information, not the representation. We then analyze the real models. The single-cell foundation model scGPT encodes biological variability only partially. We also revisit four influential LLM probing studies, which claim that models represent geography, the state of an Othello board, truth, and the demographics of their users. Against a floor computed from the input text alone, some of these claims hold, while others are largely explained by the text itself.

The development of state-of-the-art systems in different applied areas of machine learning (ML) is driven by benchmarks, which have shaped the paradigm of evaluating generalisation capabilities from multiple perspectives. Although the paradigm is shifting towards more fine-grained evaluation across diverse tasks, the delicate question of how to aggregate the performances has received particular interest in the community. In general, benchmarks follow the unspoken utilitarian principles, where the systems are ranked based on their mean average score over task-specific metrics. Such aggregation procedure has been viewed as a sub-optimal evaluation protocol, which may have created the illusion of progress. This paper proposes Vote'n'Rank, a framework for ranking systems in multi-task benchmarks under the principles of the social choice theory. We demonstrate that our approach can be efficiently utilised to draw new insights on benchmarking in several ML sub-fields and identify the best-performing systems in research and development case studies. The Vote'n'Rank's procedures are more robust than the mean average while being able to handle missing performance scores and determine conditions under which the system becomes the winner.

Multimodal Large Language Models (MLLMs) show reasoning promise, yet their visual perception is a critical bottleneck. Strikingly, MLLMs can produce correct answers even while misinterpreting crucial visual elements, masking these underlying failures. Our preliminary study on a joint perception-reasoning dataset revealed that for one leading MLLM, 29% of its correct answers to reasoning questions still exhibited visual perception errors. To systematically address this, we introduce "Do You See Me", a scalable benchmark with 1,758 images and 2,612 questions. It spans seven human-psychology inspired subtasks in 2D and 3D, featuring controllable complexity to rigorously evaluate MLLM visual skills. Our findings on 3 leading closed-source and 5 major open-source models reveal a stark deficit: humans achieve 96.49% accuracy, while top MLLMs average below 50%. This performance gap widens rapidly with increased task complexity (e.g., from 12% to 45% in the visual form constancy subtask). Further analysis into the root causes suggests that failures stem from challenges like misallocated visual attention and the instability of internal representations for fine-grained details, especially at or below encoder patch resolution. This underscores an urgent need for MLLMs with truly robust visual perception. The benchmark dataset, source code and evaluation scripts are available at https://github.com/microsoft/Do-You-See-Me.

Vision-language models (VLMs) are increasingly used to make decisions from visual inputs. We introduce FAIRLENS, a benchmark and evaluation framework for measuring both the fairness and the validity of VLM responses in three high-stakes domains: hiring, legal, and healthcare. FAIRLENS pairs real face images spanning gender, race, and age groups with closed- and open-ended questions, giving more than 100K image-question pairs per model, and evaluates responses from four complementary views: demographic parity over adverse outcome rates, soundness, demographic association over unsupported roles and statuses, and bias in free-text generation. Soundness is the central validity criterion: a response is sound when it follows the evidence stated in the question and abstains when the image cannot support an answer. Evaluating eight VLMs, we find that the primary failure is unwarranted inference rather than unequal treatment. Models routinely infer qualifications, threat, illness, or professional role from a face instead of abstaining, and the weakest model does so on 99% of the questions its input cannot answer. These failures are most severe in legal and healthcare, where recognizing insufficient evidence matters most, and disparity metrics alone would miss them: parity gaps are small in absolute terms, yet when baseline adverse rates are low the same gap means one demographic group receives adverse labels several times as often as another, and a small gap can equally reflect a model that treats every group unsafely. Bias in free-text responses is only loosely coupled to multiple-choice accuracy, so correct structured answers do not imply safe generation. FAIRLENS shows that fair high-stakes VLM behavior requires similar treatment across groups and refusal to infer high-stakes attributes from appearance, and its question suite transfers to any face corpus with demographic annotations.

In Brazil, all legal professionals must demonstrate their knowledge of the law and its application by passing the OAB exams, the national bar exams. The OAB exams therefore provide an excellent benchmark for the performance of legal information systems since passing the exam would arguably signal that the system has acquired capacity of legal reasoning comparable to that of a human lawyer. This article describes the construction of a new data set and some preliminary experiments on it, treating the problem of finding the justification for the answers to questions. The results provide a baseline performance measure against which to evaluate future improvements. We discuss the reasons to the poor performance and propose next steps.

Recent research has shown that filtering massive English web corpora into high-quality subsets significantly improves training efficiency. However, for high-resource non-English languages like German, French, or Japanese, aggressive filtering creates a strategic dilemma: should practitioners prioritize diversity by training once on large amounts of lightly filtered web data, or prioritize quality by strictly filtering for a high-quality core and repeating it over multiple epochs? We investigate this trade-off for German by constructing hierarchical quality filters applied to 500M web documents, comparing multi-epoch training on the filtered subsets against single-pass training on a diverse corpus. Our experiments across multiple model scales and token budgets show that repeating high-quality data consistently outperforms single-pass training on larger, less filtered sets. Notably, the performance gap persists even after 7 epochs. Our findings suggest that for non-English LLMs, semantic concentration through quality filtering offers a more viable path to efficient language modeling than simply maximizing unique data volume. We release our German language models (called Boldt), as well as our cleaned evaluation benchmarks to the research community. Our experiments indicate that they achieve state-of-the-art results despite training on 10-360x fewer tokens than comparable models.