Reinforcement learning with verifiable rewards (RLVR) and on-policy distillation (OPD) have emerged as two dominant methods for post-training reasoning LLMs. Prior work uses OPD’s dense token-level supervision to complement the sparse RL reward, fusing the two signals within a single step: either as a weighted-additive combination or a teacher-modulated rescaling of the RL advantage. In this paper, we show that a simple two-stage scheme, OPD-then-RL, consistently outperforms pure OPD, pure RLVR, and all such joint baselines across logic and math reasoning benchmarks. Beyond the empirical results, we further provide a systematic understanding of this through pass@k behavior, learning dynamics, and parameter updates, yielding a consistent explanation: OPD expands the student’s coverage of teacher-supported solutions and RL sharpens within that support, while jointly optimizing the two signals causes them to interfere. To provide a practical recipe, we find that the OPD validation score is the key signal for when to switch to RL, and that OPD is a better cold start for RL than SFT. Together, our results establish OPD-then-RL as a simple yet strong way to combine the two methods, turning two entangled signals into complementary stages.

AI agents may assist with emails and documents, but can they reliably complete everyday online workflows on real websites? Everyday online tasks offer a realistic yet unsolved testbed for evaluating next-generation AI agents. To this end, we introduce ClawBench, an evaluation framework comprising 153 everyday online tasks common in daily life and work, spanning 144 platforms across 15 categories, from completing purchases and booking appointments to submitting job applications. These tasks require capabilities beyond existing benchmarks, such as obtaining relevant information from user-provided documents, navigating multi-step workflows across diverse platforms, and write-heavy operations like filling in many detailed forms correctly. Unlike existing benchmarks that evaluate agents in offline sandboxes with static pages, ClawBench operates on production websites, preserving the full complexity, dynamic nature, and interaction challenges of real-world web environments. An interception layer captures and blocks the final submission request, ensuring safe evaluation without real-world side effects. Our evaluations of 8 frontier models show that both proprietary and open-source models complete only a small portion of these tasks. For example, Claude Sonnet 4.6 achieves only 33.3%, which exposes gaps in current AI agents. Progress on ClawBench brings us closer to AI agents that can function as general-purpose assistants.

Multimodal large language models (MLLMs) have recently shown strong potential as point-wise rerankers by directly modeling query–document relevance through next-token prediction. However, point-wise reranking suffers from substantial repeated computation across query–document pairs, while the causal structure of transformers allows only prefix segments to be reused via pre-caching. To address the misalignment of existing query-first and document-first formats with both VQA-style prompting and computation-aware reuse, we propose a vision-first formulation that improves both cache reuse efficiency and reranking performance. However, the remaining cost is still considerable and stems from three main sources: (1) *model depth*, for which we reduce active parameters via early exit; (2) *cross-segment attention*, which we restrict to a narrow interaction band across a few layers; and (3) *visual tokens*, where we reduce the number of tokens via embedder-guided pruning. Together, these designs form *miniReranker*, which reduces reranking runtime to <1% of the dense implementation under high-reuse settings for a single query, while preserving >96% of the dense model performance.

Agentic repository-level code understanding is essential for automating complex software engineering tasks, yet the field lacks reliable benchmarks. Existing evaluations often overlook the long tail topics and rely on popular repositories where Large Language Models (LLMs) can cheat via memorized knowledge. To address this, we introduce SWE-QA-Pro, a benchmark constructed from diverse, long-tail repositories with executable environments. We enforce topical balance via issue-driven clustering to cover under-represented task types and apply a rigorous difficulty calibration process: questions solvable by direct-answer baselines are filtered out. This results in a dataset where agentic workflows significantly outperform direct answering (e.g., a ~13-point gap for Claude Sonnet 4.5), confirming the necessity of agentic codebase exploration. Furthermore, to tackle the scarcity of training data for such complex behaviors, we propose a scalable synthetic data pipeline that powers a two-stage training recipe: Supervised Fine-Tuning (SFT) followed by Reinforcement Learning from AI Feedback (RLAIF). This approach allows small open models to learn efficient tool usage and reasoning. Empirically, a Qwen3-8B model trained with our recipe surpasses GPT-4o by 2.3 points on SWE-QA-Pro and substantially narrows the gap to state-of-the-art proprietary models, demonstrating both the validity of our evaluation and the effectiveness of our agentic training workflow.

The rapid growth of scientific literature calls for automated methods to assess and predict research impact.Prior work has largely focused on citation-based metrics, leaving limited evaluation of models’ capability to reason about other impact dimensions.To this end, we introduce SciImpact, a large-scale, multi-dimensional benchmark for scientific impact prediction spanning 19 fields.SciImpact captures various forms of scientific influence, ranging from citation counts to award recognition, media attention, patent reference, and artifact adoption, by integrating heterogeneous data sources and targeted web crawling.It comprises 215,928 contrastive paper pairs reflecting meaningful impact differences in both short- (e.g., Best Paper Award) and long-term settings (e.g., Nobel Prize).We evaluate 11 widely used large language models (LLMs) on SciImpact.Results show that off-the-shelf models show substantial variability across dimensions and fields, while multi-task supervised fine-tuning consistently enables smaller LLMs (e.g., 4B) to markedly outperform much larger models (e.g., 30B) and surpass powerful closed-source LLMs (e.g., o4-mini).These results establish SciImpact as a challenging benchmark and demonstrate its value for multi-dimensional, multi-field scientific impact prediction.Our project homepage is https://flypig23.github.io/sciimpact-homepage/.

Training deep research agents requires long-horizon trajectories that interleave search, evidence aggregation, and multi-step reasoning. However, existing data collection pipelines typically rely on proprietary web APIs, making large-scale trajectory synthesis costly, unstable, and difficult to reproduce. We present OpenResearcher, a reproducible pipeline that decouples one-time corpus bootstrapping from multi-turn trajectory synthesis and executes the search-and-browse loop entirely offline using three explicit browser primitives: search, open, and find, over a 15M-document corpus. Using GPT-OSS-120B as the teacher model, we synthesize over 97K trajectories, including a substantial long-horizon tail with 100+ tool calls. Supervised fine-tuning a 30B-A3B backbone on these trajectories achieves 54.8% accuracy on BrowseComp-Plus, an absolute improvement of 34.0 percentage points over the base model, while remaining competitive on BrowseComp, GAIA, and xbench-DeepSearch. Because the environment is offline and fully instrumented, it also enables controlled analysis, where our study reveals practical insights into deep research pipeline design, including data filtering strategies, agent configuration choices, and how retrieval success relates to final answer accuracy.

Standard on-policy distillation (OPD) for large language models estimates the reverse-KL objective using student-sampled tokens, yielding an unbiased single-sample Monte Carlo estimator that avoids vocabulary-wide computation. However, we show that this estimator suffers from severe training pathologies in practice: sample inefficiency, unstable generation dynamics, and a substantial performance gap compared to exact full-vocabulary OPD. Reward-level diagnosis traces these pathologies to the log-ratio reward, which is unbounded by construction, producing extremely high-variance gradients concentrated at early positions and persisting throughout training; standard post-hoc scaling fail as they operate only after this distortion occurs. To solve this problem, we propose *PowerOPD*: a family of natively bounded, sign-consistent rewards from the Box-Cox power transformation, parameterized by 𝛼 > 0, of which the log-ratio is the degenerate 𝛼 → 0 limit. Across six mathematical reasoning benchmarks and four Qwen3 teacher–student pairs, PowerOPD achieves benchmark-averaged Avg@8/Pass@8 gains of up to +6.37/+5.71 over vanilla OPD, +3.01/+3.54 over post-hoc stabilization, and +2.59/+8.90 over full-vocabulary OPD, while reducing wall-clock time by 59.2% and peak GPU memory by 23.1%. Larger 𝛼 generally improves accuracy, consistently shortens responses, and keeps gradient norms more than 3,000× smaller than vanilla OPD.

Recent studies suggest that video generative models can exhibit certain forms of zero-shot visual reasoning through generated frames. Yet reliable evaluation remains challenging: benchmarks should adopt inputs aligned with the visual priors of current video models, require valid evolving processes rather than only plausible final states, and calibrate task difficulty to remain challenging yet partly feasible. To this end, we introduce VGI-Bench, containing 27 tasks and 810 instances, organized by a two-level taxonomy of domains and skill tags for evaluating the visual reasoning capabilities of video generative models. Our evaluations show that current generative systems can solve a subset of visually grounded reasoning tasks, but remain far from reliable, with even the strongest model, Seedance 2.0, achieving only 51.0% under our evaluation criteria. Our findings further reveal the input sensitivity, skill-bounded transfer from synthetic fine-tuning, and limited denoising-time self-correction, where later steps mainly refine early hypotheses rather than correct reasoning errors. We hope VGI-Bench will help stimulate the development of next-generation video generative models. Data and code are available at [our project page](https://hexuan21.github.io/VGI-Bench/).

Existing benchmarks for Deep Research Agents (DRAs) treat report generation as a single-shot writing task, which fundamentally diverges from how human researchers iteratively draft and revise reports via self-reflection or peer feedback. Whether DRAs can reliably revise reports with user feedback remains unexplored. We introduce Mr Dre, an evaluation suite that establishes multi-turn report revision as a new axis. Mr Dre consists of (1) a unified long-form report evaluation protocol spanning comprehensiveness, factuality, and presentation, and (2) a human-verified feedback simulation pipeline for systematic multi-turn revision evaluation. Our analysis of five diverse DRAs reveals a critical limitation: while agents can address most user feedback, they also regress on 16–27% of previously covered content and citation quality. Over multiple revision turns, even the best-performing agents leave significant headroom, as they continue to disrupt content outside the feedback’s scope and fail to preserve earlier edits. We further show that these issues are not easily resolvable through inference-time fixes such as prompt engineering and a dedicated sub-agent for revision.

Deep search agents that combine large language models with retrieval tools excel at complex, multi-hop queries. Yet, existing benchmarks such as BrowseComp rely on black-box web search APIs, facing key limitations. (1) Fairness: for agents, dynamic and opaque web APIs hinder reproducibility and fair comparisons across agents. (2) Disentanglement: for retrieval, the lack of a fixed document corpus makes it impossible to isolate retriever contributions from end-to-end search agent accuracy. We introduce BrowseComp-Plus, a benchmark derived from BrowseComp that employs a fixed, human-verified corpus, enabling controlled retrieval for deep search agents. BrowseComp-Plus clearly distinguishes agent performance: with a BM25 retriever, the open-source Search-R1 achieves 3.86% accuracy, while GPT-5 achieves 55.9%. Additionally, BrowseComp-Plus makes retrieval gains explicit: pairing GPT-5 with Qwen3-Embedding-8B retriever further improves accuracy to 70.1% while reducing search calls. Overall, BrowseComp-Plus provides a fair and disentangled testbed, advancing both deep search agent evaluation and retrieval research for agentic search. Code and data can be found at: https://texttron.github.io/BrowseComp-Plus/

Large language models (LLMs) often struggle with visualization tasks like plotting diagrams, charts, where success depends on both code correctness and visual semantics. Existing instruction-tuning datasets lack execution-grounded supervision and offer limited support for iterative code correction, resulting in fragile and unreliable plot generation. We present VisCode-200K, a large-scale instruction tuning dataset for Python-based visualization and self-correction. It contains over 200K examples from two sources: (1) validated plotting code from open-source repositories, paired with natural language instructions and rendered plots; and (2) 45K multi-turn correction dialogues from Code-Feedback, enabling models to revise faulty code using runtime feedback. We fine-tune Qwen2.5-Coder-Instruct on VisCode-200K to create VisCoder, and evaluate it on PandasPlotBench. VisCoder significantly outperforms strong open-source baselines and approaches the performance of proprietary models like GPT-4o-mini. We further adopt a self-debug evaluation protocol to assess iterative repair, demonstrating the benefits of feedback-driven learning for executable, visually accurate code generation.

Vision-Language Models have made significant progress on many perception-focused tasks. However, their progress on reasoning-focused tasks remains limited due to the lack of high-quality and diverse training data. In this work, we aim to address the scarcity of reasoning-focused multimodal datasets. We propose VisualWebInstruct, a novel approach that leverages search engines to create a diverse and high-quality dataset spanning multiple disciplines, including mathematics, physics, finance, and chemistry, etc. Starting with a meticulously selected set of 30,000 seed images, we employ Google Image Search to identify websites containing similar images. We collect and process HTML data from over 700K unique URLs. Through a pipeline of content extraction, filtering, and synthesis, we construct a dataset of approximately 900K question-answer (QA) pairs, with 40% consisting of visual QA pairs and the remaining comprising text-based QA pairs. Models fine-tuned on VisualWebInstruct demonstrate significant performance improvements: (1) fine-tuning on Llava-OV results in 10-20 absolute points improvement across benchmarks, and (2) fine-tuning from MAmmoTH-VL yields a 5 absolute points gain across benchmarks. Our best model, MAmmoTH-VL2, achieves the best known performance with SFT without RL within the 10B parameter class on MMMU-Pro (40.7), MathVerse (42.6), and DynaMath (55.7). These results highlight the effectiveness of our dataset in enhancing the reasoning capabilities of vision-language models for complex multimodal tasks.

Critique Fine-Tuning (CFT) has recently emerged as a promising paradigm for unlocking the reasoning capabilities of large language models (LLMs). In this work, we introduce one-shot CFT, a highly compute-efficient approach that leverages critique data generated from a single math problem. Remarkably, this method yields significant gains in reasoning accuracy, surpassing one-shot RLVR (Reinforcement Learning with Verifiable Reward) while requiring 15 to 20 times less compute. Given one math problem, we first prompt a set of diverse small models to produce candidate solutions, then use frontier models such as GPT-4.1 to generate high-quality critiques of these responses. We fine-tune Qwen and Llama family models ranging from 1.5B to 14B parameters with CFT. With just 5 GPU hours, our models achieve up to a 16 percent absolute improvement in average accuracy across six mathematical reasoning benchmarks (for example, Qwen2.5-Math-7B improves from 26 percent to 42 percent). Furthermore, ablation studies reveal the robustness of one-shot CFT across different prompt problems. Our findings suggest an extremely compute-efficient approach to unleash the reasoning potential of LLMs.

Most progress in recent coder models has been driven by supervised fine-tuning (SFT), while the potential of reinforcement learning (RL) remains largely unexplored, primarily due to the lack of reliable reward data/model in the code domain. In this paper, we address this challenge by leveraging automated large-scale test-case synthesis to enhance code model training. Specifically, we design a pipeline that generates extensive (question, test-cases) pairs from existing code data. Using these test cases, we construct preference pairs based on pass rates over sampled programs to train reward models with Bradley-Terry loss. It shows an average of 10-point improvement for Llama-3.1-8B-Ins and 5-point improvement for Qwen2.5-Coder-7B-Ins through best-of-32 sampling, making the 7B model on par with 236B DeepSeek-V2.5. Furthermore, we conduct reinforcement learning with both reward models and test-case pass rewards, leading to consistent improvements across HumanEval, MBPP, BigCodeBench, and LiveCodeBench (V4). Notably, we follow the R1-style training to start from Qwen2.5-Coder-base directly and show that our RL training can improve model on HumanEval-plus by over 25% and MBPP-plus by 6% for merely 80 optimization steps. We believe our results highlight the huge potential of reinforcement learning in coder models.

Retrieval-Augmented Generation (RAG) significantly improved the ability of Large Language Models (LLMs) to solve knowledge-intensive tasks. While existing research seeks to enhance RAG performance by retrieving higher-quality documents or designing RAG-specific LLMs, the internal mechanisms within LLMs that contribute to RAG’s effectiveness remain underexplored. In this paper, we aim to investigate these internal mechanisms within the popular Mixture-of-Expert (MoE)-based LLMs and demonstrate how to improve RAG by examining expert activations in these LLMs. Our controlled experiments reveal that several core groups of experts are primarily responsible for RAG-related behaviors. The activation of these core experts can signify the model’s inclination towards external/internal knowledge and adjust its behavior. For instance, we identify core experts that can (1) indicate the sufficiency of the model’s internal knowledge, (2) assess the quality of retrieved documents, and (3) enhance the model’s ability to utilize context. Based on these findings, we propose several strategies to enhance RAG’s efficiency and effectiveness through expert activation. Experimental results across various datasets and MoE LLMs show the effectiveness of our method.

High-quality datasets are significant to the development of dialogue models. However, most existing datasets for open-domain dialogue modeling are limited to a single language. The absence of multilingual open-domain dialog datasets not only limits the research on multilingual or cross-lingual transfer learning, but also hinders the development of robust open-domain dialog systems that can be deployed in other parts of the world. In this paper, we provide a multilingual parallel open-domain dialog dataset, XDailyDialog, to enable researchers to explore the challenging task of multilingual and cross-lingual open-domain dialog. XDailyDialog includes 13K dialogues aligned across 4 languages (52K dialogues and 410K utterances in total). We then propose a dialog generation model, kNN-Chat, which has a novel kNN-search mechanism to support unified response retrieval for monolingual, multilingual, and cross-lingual dialogue. Experiment results show the effectiveness of this framework. We will make XDailyDialog and kNN-Chat publicly available soon.

Recent work has demonstrated that pre-trained language models (PLMs) are zero-shot learners. However, most existing zero-shot methods involve heavy human engineering or complicated self-training pipelines, hindering their application to new situations. In this work, we show that zero-shot text classification can be improved simply by clustering texts in the embedding spaces of PLMs. Specifically, we fit the unlabeled texts with a Bayesian Gaussian Mixture Model after initializing cluster positions and shapes using class names. Despite its simplicity, this approach achieves superior or comparable performance on both topic and sentiment classification datasets and outperforms prior works significantly on unbalanced datasets. We further explore the applicability of our clustering approach by evaluating it on 14 datasets with more diverse topics, text lengths, and numbers of classes. Our approach achieves an average of 20% absolute improvement over prompt-based zero-shot learning. Finally, we compare different PLM embedding spaces and find that texts are well-clustered by topics even if the PLM is not explicitly pre-trained to generate meaningful sentence embeddings. This work indicates that PLM embeddings can categorize texts without task-specific fine-tuning, thus providing a new way to analyze and utilize their knowledge and zero-shot learning ability.