Humans acquire knowledge through three cognitive stages: perceiving information, comprehending knowledge, and adapting knowledge to solve novel problems. Videos serve as an effective medium for knowledge acquisition, facilitating a natural progression through these learning stages. However, existing video benchmarks fail to evaluate the knowledge acquisition capabilities of Large Multimodal Models (LMMs). To address this gap, we introduce Video-MMMU, a multi-modal, multi-discipline, multi-track benchmark that evaluates LMMs’ ability to acquire knowledge from college-level, educational videos. Video-MMMU features a collection of 300 videos and 900 human-annotated questions across six disciplines, evaluating knowledge acquisition through stage-aligned question-answer pairs: Perception, Comprehension, and Adaptation. Beyond measuring final accuracy, Video-MMMU proposes the performance gain metric that quantifies an LMM’s learning gain from video, shifting the focus of evaluation from absolute performance to learning efficiency. Our evaluation reveals a substantial gap between human learners and current LMMs, highlighting the need to improve models’ ability to learn and adapt knowledge from video content.

Current benchmarks for occupational AI agents are scoped primarily by economic values, telling a replacement story. We introduce JobBench, which evaluates AI agents on the workflows that experts identify as high-priority for delegation, empowering humans based on their needs instead of replacing them with GDP value. JobBench covers 130 agentic tasks across 35 occupations. Each task is packaged as a workspace of heterogeneous reference files, requiring the agent to reason through the cluttered information streams of real professional work. Outputs are graded by a fact-anchored chain of rubrics, averaging 35.6 binary criteria per task. We evaluate 36 models; the strongest, Claude Opus 4.7 under Claude Code, reaches only 45.9

Real-world language agents must handle complex, multi-step workflows across diverse applications. For instance, an agent may manage emails by coordinating with calendars and file systems, or monitor a production database like BigQuery to detect anomalies and generate reports following a standard operating manual. However, existing language agent benchmarks often focus on narrow domains or simplified tasks that lack the diversity, realism, and long-horizon complexity required to evaluate agents' real-world performance. To address this gap, we introduce the Tool Decathlon (dubbed as Toolathlon), a benchmark for language agents offering diverse applications and tools, realistic environment setup, and reliable execution-based evaluation. Toolathlon spans 32 software applications and 604 tools, ranging from everyday platforms such as Google Calendar and Notion to professional applications like WooCommerce, Kubernetes, and BigQuery. Most of the tools are based on a high-quality set of Model Context Protocol (MCP) servers that we may have revised or implemented ourselves. Unlike prior works, which primarily ensure functional realism but offer limited environment state diversity, we provide realistic initial environment states from real software, such as multiple Canvas courses each with dozens of students or real-world financial spreadsheets. The Toolathlon benchmark includes 108 manually sourced or crafted tasks in total, requiring interacting with multiple applications over ~20 interaction turns on average to complete. Each task is strictly verifiable through dedicated evaluation scripts. Comprehensive evaluation of state-of-the-art models highlights their significant shortcomings in performing real-world, long-horizon tasks: the best-performing model, Claude-4-Sonnet, achieves only a 29.9% success rate with 28 tool calling turns on average, while the top open-weights model DeepSeek-V3.1 reaches 13.9%. We expect Toolathlon to drive the development of more capable language agents for real-world, long-horizon task execution.

Fine-tuning large language models (LLMs) is intended to improve their reasoning capabilities, yet we uncover a counterintuitive effect: models often forget how to solve problems they previously answered correctly during training. We term this phenomenon temporal forgetting and show that it is widespread across model sizes, fine-tuning methods (both Reinforcement Learning and Supervised Fine-Tuning), and multiple reasoning benchmarks. To address this gap, we introduce Temporal Sampling, a simple decoding strategy that draws outputs from multiple checkpoints along the training trajectory. This approach recovers forgotten solutions without retraining or ensembling, and leads to substantial improvements in reasoning performance, gains from 4 to 19 points in Pass@k and consistent gains in Majority@k across several benchmarks. We further extend our method to LoRA-adapted models, demonstrating that storing only adapter weights across checkpoints achieves similar benefits with minimal storage cost. By leveraging the temporal diversity inherent in training, Temporal Sampling offers a practical, compute-efficient way to surface hidden reasoning ability and rethink how we evaluate LLMs.

As language model (LM) outputs get more and more natural, it is becoming more difficult than ever to evaluate their quality. Simultaneously, increasing LMs' "thinking" time through scaling test-time compute has proven an effective technique to solve challenging problems in domains such as math and code. This raises a natural question: can an LM's evaluation capability also be improved by spending more test-time compute? To answer this, we investigate employing reasoning models-LMs that natively generate long chain-of-thought reasoning-as evaluators. Specifically, we examine methods to leverage more test-time compute by (1) using reasoning models, and (2) prompting these models to evaluate not only the response as a whole (i.e., outcome evaluation) but also assess each step in the response separately (i.e., process evaluation). In experiments, we observe that the evaluator's performance improves monotonically when generating more reasoning tokens, similar to the trends observed in LM-based generation. Furthermore, we use these more accurate evaluators to rerank multiple generations, and demonstrate that spending more compute at evaluation time can be as effective as using more compute at generation time in improving an LM's problem-solving capability.

Math reasoning has become the poster child of progress in large language models (LLMs), with new models rapidly surpassing human-level performance on benchmarks like MATH and AIME. But as math leaderboards improve week by week, it is worth asking: do these gains reflect broader problem-solving ability or just narrow overfitting? To answer this question, we evaluate over 20 open-weight reasoning-tuned models across a broad suite of tasks, including math, scientific QA, agent planning, coding, and standard instruction-following. We surprisingly find that most models that succeed in math fail to transfer their gains to other domains. To rigorously study this phenomenon, we conduct controlled experiments on Qwen3-14B models using math-only data but different tuning methods. We find that reinforcement learning (RL)-tuned models generalize well across domains, while supervised fine-tuning (SFT)-tuned models often forget general capabilities. Latent-space representation and token-space distribution shift analyses reveal that SFT induces substantial representation and output drift, while RL preserves general-domain structure. Our results suggest a need to rethink standard post-training recipes, particularly the reliance on SFT-distilled data for advancing reasoning models.

With the advancement of AI capabilities, AI reviewers are beginning to be deployed in scientific peer review, yet their capability and credibility remain in question: many scientists simply view them as probabilistic systems without the expertise to evaluate research, while other researchers are more optimistic about their readiness without concrete evidence. Understanding what AI reviewers do well, where they fall short, and what challenges remain is essential. However, existing evaluations of AI reviewers have focused on whether their verdicts match human verdicts (e.g., score alignment, acceptance prediction), which is insufficient to characterize their capabilities and limits. In this paper, we close this gap through a large-scale expert annotation study, in which 45 domain scientists in Physical, Biological, and Health Sciences spent 469 hours rating 2,960 individual criticisms (each targeting one specific aspect of a paper) from human-written and AI-generated reviews of 82 Nature-family papers on correctness, significance, and sufficiency of evidence. On a composite of all three dimensions, a reviewing agent powered by GPT-5.2 scores above each paper's top-rated human reviewer (60.0

Evaluating LLM agents on benchmarks like SWE-Bench and GAIA can be expensive, time-consuming, and requires complex infrastructure. A single evaluation can cost thousands of dollars and take days to complete. In contrast, non-agentic LLM benchmarks that test individual capabilities (e.g., reasoning, code generation) are fast and cheap to run. In this paper, we investigate whether performance on expensive agentic benchmarks can be accurately predicted by the performance on a small, carefully selected subset of atomic evaluation instances. We introduce PACE, a framework that constructs proxy benchmarks by selecting instances from existing non-agentic evaluations whose aggregate scores most reliably predict model performances on agentic benchmarks. Given a pool of candidate instances spanning atomic capabilities, PACE fits a regression that maps a model's scores on a compact subset of source instances to its score on the target agentic benchmark. The subset itself is curated by combining two complementary instance-selection strategies, target-relevance local selection and globally informative global selection. We apply PACE to the 4 target agentic benchmarks in this paper, which yields PACE-Bench, the concrete proxy benchmark that we evaluate in the paper. Experiments across 14 models, 4 agentic benchmarks, and 19 non-agentic benchmarks show that PACE-Bench predicts agentic scores with leave-one-out cross-validation (LOOCV) mean absolute error (MAE) under 4

Current multimodal benchmarks often conflate reasoning with domain-specific knowledge, making it difficult to isolate and evaluate general reasoning abilities in non-expert settings. To address this, we introduce VisualPuzzles, a benchmark that targets visual reasoning while deliberately minimizing reliance on specialized knowledge. VisualPuzzles consists of diverse questions spanning five categories: algorithmic, analogical, deductive, inductive, and spatial reasoning. One major source of our questions is manually translated logical reasoning questions from the Chinese Civil Service Examination. Experiments show that VisualPuzzles requires significantly less intensive domain-specific knowledge and more complex reasoning compared to benchmarks like MMMU, enabling us to better evaluate genuine multimodal reasoning. Evaluations show that state-of-the-art multimodal large language models consistently lag behind human performance on VisualPuzzles, and that strong performance on knowledge-intensive benchmarks does not necessarily translate to success on reasoning-focused, knowledge-light tasks. Additionally, reasoning enhancements such as scaling up inference compute (with "thinking" modes) yield inconsistent gains across models and task types, and we observe no clear correlation between model size and performance. We also found that models exhibit different reasoning and answering patterns on VisualPuzzles compared to benchmarks with heavier emphasis on knowledge. VisualPuzzles offers a clearer lens through which to evaluate reasoning capabilities beyond factual recall and domain knowledge.

Large language models (LLMs) have recently enabled coding agents capable of generating, executing, and revising visualization code. However, existing models often fail in practical workflows due to limited language coverage, unreliable execution, and lack of iterative correction mechanisms. Progress has been constrained by narrow datasets and benchmarks that emphasize single-round generation and single-language tasks. To address these challenges, we introduce three complementary resources for advancing visualization coding agents. **VisCode-Multi-679K** is a large-scale, supervised dataset containing 679K validated and executable visualization samples with multi-turn correction dialogues across 12 programming languages. **VisPlotBench** is a benchmark for systematic evaluation, featuring executable tasks, rendered outputs, and protocols for both initial generation and multi-round self-debug. Finally, we present **VisCoder2**, a family of multi-language visualization models trained on VisCode-Multi-679K. Experiments show that VisCoder2 significantly outperforms strong open-source baselines and approaches the performance of proprietary models like GPT-4.1, with further gains from iterative self-debug, reaching **82.4%** overall execution pass rate at the 32B scale, particularly in symbolic or compiler-dependent languages.

Perceiving visual semantics embedded within consecutive characters is a crucial yet under-explored capability for both Large Language Models (LLMs) and Multi-modal Large Language Models (MLLMs). In this work, we select ASCII art as a representative artifact. It depicts concepts through careful arrangement of characters, which can be formulated in both text and image modalities. We frame the problem as a recognition task, and construct a novel benchmark, ASCIIEval. It covers over 3K samples with an elaborate categorization tree, along with a training set for further enhancement. Encompassing a comprehensive analysis of tens of models through different input modalities, our benchmark demonstrate its multi-faceted diagnostic power. Given textual input, language models shows their visual perception ability on ASCII art concepts. Proprietary models achieve over 70\% accuracy on certain categories, with GPT-5 topping the rank. For image inputs, we reveal that open-source MLLMs suffer from a trade-off between fine-grained text recognition and collective visual perception. They exhibit limited generalization ability to this special kind of arts, leading to the dramatic gap of over 20.01\% accuracy compared with their proprietary counterparts. Another critical finding is that model performance is sensitive to the length of the ASCII art, with this sensitivity varying across input modalities. Unfortunately, none of the models could successfully benefit from the simultaneous provision of both modalities, highlighting the need for more flexible modality-fusion approaches. Besides, we also introduce approaches for further enhancement and discuss future directions. Resources are available at https://anonymous.4open.science/r/VisionInText-9EE5.

Long chain-of-thought (CoT) is an essential ingredient in effective usage of modern large language models, but our understanding of the reasoning strategies underlying these capabilities remains limited. While some prior works have attempted to categorize CoTs using predefined strategy types, such approaches are constrained by human intuition and fail to capture the full diversity of model behaviors. In this work, we introduce the CoT Encyclopedia, a bottom-up framework for analyzing and steering model reasoning. Our method automatically extracts diverse reasoning criteria from model-generated CoTs, embeds them into a semantic space, clusters them into representative categories, and derives contrastive rubrics to interpret reasoning behavior. Human evaluations show that this framework produces more interpretable and comprehensive analyses than existing methods. Moreover, we demonstrate that this understanding enables performance gains: we can predict which strategy a model is likely to use and guide it toward more effective alternatives. Finally, we provide practical insights, such as that training data format (e.g., free-form vs. multiple-choice) has a far greater impact on reasoning behavior than data domain, underscoring the importance of format-aware model design.

Recent reinforcement learning (RL) techniques have yielded impressive reasoning improvements in language models, yet it remains unclear whether post-training truly extends a model's reasoning ability beyond what it acquires during pre-training. A central challenge is the lack of control in modern training pipelines: large-scale pre-training corpora are opaque, mid-training is often underexamined, and RL objectives interact with unknown prior knowledge in complex ways. To resolve this ambiguity, we develop a fully controlled experimental framework that isolates the causal contributions of pre-training, mid-training, and RL-based post-training. Our approach employs synthetic reasoning tasks with explicit atomic operations, parseable step-by-step reasoning traces, and systematic manipulation of training distributions. We evaluate models along two axes: extrapolative generalization to more complex compositions and contextual generalization across surface contexts. Using this framework, we reconcile competing views on RL's effectiveness. We show that: 1) RL produces true capability gains (pass@128) only when pre-training leaves sufficient headroom and when RL data target the model's edge of competence, tasks at the boundary that are difficult but not yet out of reach. 2) Contextual generalization requires minimal yet sufficient pre-training exposure, after which RL can reliably transfer. 3) Mid-training significantly enhances performance under fixed compute compared with RL only, demonstrating its central but underexplored role in training pipelines. 4) Process-level rewards reduce reward hacking and improve reasoning fidelity. Together, these results clarify the interplay between pre-training, mid-training, and RL, offering a foundation for understanding and improving reasoning LM training strategies.

Electrocardiograms (ECGs) are essential, non-invasive diagnostic tools for assessing cardiac conditions. Existing methods often have limited generalizability, focus on narrow condition sets, and rely on raw physiological signals, which may be unavailable in resource-limited settings where only printed or digital ECG images are accessible. Recent advances in multimodal large language models (MLLMs) offer new opportunities, yet ECG image interpretation remains challenging due to the lack of instruction-tuning data and standardized benchmarks. To address these gaps, we introduce ECGInstruct, the first large-scale ECG image instruction-tuning dataset with over one million samples, covering diverse tasks including feature recognition, rhythm analysis, morphology assessment, and clinical report generation. We develop PULSE, a fully open-source MLLM for ECG image interpretation trained on ECGInstruct. We further curate ECGBench, a human expert-developed benchmark spanning four core ECG interpretation tasks across nine datasets, incorporating both synthesized and real-world ECG images to enable clinically realistic evaluation. Our experiments demonstrate that PULSEestablishes a new state of the art, outperforming general-purpose MLLMs by 21% to 33% in average accuracy. These results highlight the potential of PULSEto improve ECG image interpretation in clinical practice. All code, data and models are available at https://aimedlab.github.io/PULSE/.

Unlearning in large language models (LLMs) aims to remove specified data, but its efficacy is typically assessed with task-level metrics like accuracy and perplexity. We demonstrate that these metrics are often misleading, as models can appear to forget while their original behavior is easily restored through minimal fine-tuning. This phenomenon of \emph{reversibility} suggests that information is merely suppressed, not genuinely erased. To address this critical evaluation gap, we introduce a \emph{representation-level analysis framework}. Our toolkit comprises PCA similarity and shift, centered kernel alignment (CKA), and Fisher information, complemented by a summary metric, the mean PCA distance, to measure representational drift. Applying this framework across six unlearning methods, three data domains, and multiple LLMs, we identify four distinct forgetting regimes based on their \emph{reversibility} and \emph{catastrophicity}. We compare recovery strategies and show that relearning efficiency relies on the data source. We also find that irreversible, non-catastrophic forgetting is exceptionally challenging. By probing unlearning limits, we identify a case of seemingly irreversible, targeted forgetting, offering insights for more robust erasure algorithms. Overall, our findings expose a gap in current evaluation and establish a representation-level foundation for trustworthy unlearning.

Models are no longer merely trained and then used, but generate experience during use from which future behavior should improve. A deployed assistant may learn user preferences, a coding agent may encounter recurring repository conventions, and an interactive agentic system may receive corrections, tool outcomes, or environmental feedback. The question is therefore not only how to train a capable model, but how a model should keep, update, and reuse what happens during use. We use test-time learning (TTL) to describe a deployment-time learning paradigm that complements pre-training and post-training. TTL occurs when experience encountered during use updates adaptive state that causally shapes a subsequent decision, forming a deployment-time write-read loop. This formulation provides a unifying lens for a fragmented body of work on memory, adaptation, model editing, and self-evolving agents. We formalize the loop and organize qualifying forms of TTL into context-level, representation-level, and parameter-level TTL, according to where adaptive state re-enters the system. This taxonomy makes lines of work that grew up in separate communities comparable as variants of the same problem: how to turn use into durable learning. We close by reviewing how the causal contribution of learned state should be evaluated, how learning accumulates and remains stable over time, and the challenges facing trustworthy learning during deployment.

Open-source multimodal large language models (MLLMs) have shown significant potential in a broad range of tasks. However, their reasoning capabilities remain constrained by existing instruction-tuning datasets, which were predominately repurposed from academic datasets such as VQA, AI2D, and ChartQA. These datasets target simplistic tasks, and only provide phrase-level answers without any intermediate rationales.To address these challenges, we introduce a scalable and cost-effective method to construct a large-scale multimodal instruction-tuning dataset with rich intermediate rationales designed to elicit CoT reasoning. Using only open models, we create a dataset containing 12M instruction-response pairs to cover diverse reasoning-intensive tasks.Experiments demonstrate that training MLLMs on our dataset not only significantly improves reasoning capabilities, achieving state-of-the-art performance on benchmarks such as MathVerse (+8.1%), MMMU-Pro (+7%), and MuirBench (+13.3%), but also gains improvements of up to 4% on non-reasoning-based benchmarks.

Large language models (LLMs) excel in complex reasoning tasks, and distilling their reasoning capabilities into smaller models has shown promise. However, we uncover an interesting phenomenon, which we term the Small Model Learnability Gap: small models ($\leq$3B parameters) do not consistently benefit from long chain-of-thought (CoT) reasoning or distillation from larger models. Instead, they perform better when fine-tuned on shorter, simpler reasoning chains that better align with their intrinsic learning capacity. To address this, we propose Mix Distillation, a simple yet effective strategy that balances reasoning complexity by combining long and short CoT examples or reasoning from both larger and smaller models. Our experiments demonstrate that Mix Distillation significantly improves small model reasoning performance compared to training on either data alone. These findings highlight the limitations of direct strong model distillation and underscore the importance of adapting reasoning complexity for effective reasoning capability transfer.

Accelerating research on Large Multimodal Models (LMMs) in non-English languages is crucial for enhancing user experiences across broader populations. In this paper, we introduce JMMMU (Japanese MMMU), the first large-scale Japanese benchmark designed to evaluate LMMs on expert-level tasks based on the Japanese cultural context. To facilitate comprehensive culture-aware evaluation, JMMMU features two complementary subsets: (i) culture-agnostic (CA) subset, where the culture-independent subjects (e.g., Math) are selected and translated into Japanese, enabling one-to-one comparison with its English counterpart MMMU; and (ii) culture-specific (CS) subset, comprising newly crafted subjects that reflect Japanese cultural context. Using the CA subset, we observe performance drop in many LMMs when evaluated in Japanese, which is purely attributable to language variation. Using the CS subset, we reveal their inadequate Japanese cultural understanding. Further, by combining both subsets, we identify that some LMMs perform well on the CA subset but not on the CS subset, exposing a shallow understanding of the Japanese language that lacks depth in cultural understanding. We hope this work will not only help advance LMM performance in Japanese but also serve as a guideline to create high-standard, culturally diverse benchmarks for multilingual LMM development.