Strong large language models (LLMs) often achieve high scores on many benchmarks, yet can fail on closely related, sometimes easier, problems near their apparent capability boundary, a phenomenon known as jagged intelligence that reveals reasoning weaknesses masked by aggregate scores. Existing evaluations of jagged intelligence rely on manually designed benchmark-adjacent rewrites, which are hard to scale and quickly saturate as models improve. To address this challenge, we propose Automated Weakness Exploration (AWE), a fully automated framework for constructing and maintaining jaggedness evaluations. Using AWE, we construct JaggedBench with 1,142 verified capability-adjacent variants and evaluate 24 state-of-the-art models. On the primary evaluation set AIME 2025, benchmark-adjacent rewrites induce an average OG-to-RW relative drop of approximately 25% in Avg@k across 24 models, with consistent degradation observed across math, code, science, and knowledge domains. We identify 18 recurring failure modes. Our analysis shows that increasing reasoning intensity and scaling model size improve average performance and reduce failure frequency, yet the underlying failure-mode categories largely persist. Overall, these findings indicate that standard benchmark scores can present an overly optimistic view of reliability near the capability boundary. JaggedBench provides a more faithful and scalable lens for evaluating boundary robustness, enabling systematic diagnosis of boundary failures and longitudinal tracking of persistent weaknesses as models evolve. We release the benchmark, code, and subsequent updates at https://jaggedbench.github.io/.

Parallel thinking offers a promising avenue for scaling test-time compute in Large Language Models (LLMs), enabling them to explore diverse solution paths simultaneously before aggregating them into a final answer. However, coordinating the exploration and aggregation stages remains challenging, as simple aggregation techniques often incur information loss, failing to preserve the subtle, decision-relevant signals generated during exploration. To overcome this, we propose Rhombus, a parallel thinking framework that explicitly incentivizes coordination between components via end-to-end reinforcement learning. Rhombus employs multiple parallel Proposers to generate compact, decision-focused reasoning cues and a central Synthesizer to integrate them into final predictions, utilizing co-training under a shared task reward to align their interaction. Across challenging mathematical reasoning benchmarks, Rhombus improves accuracy by 6.0% over long chain-of-thought baselines while reducing wall-clock latency by 39.4% under matched token budgets. Our work demonstrates that explicit communication optimization is essential for realizing the accuracy and efficiency gains of parallel reasoning.

Large Language Models (LLMs) increasingly generate executable visual artifacts, yet established code benchmarks largely overlook the visual fidelity and interaction behavior that shape user experience. We introduce ArtifactsBench, a benchmark and automated multimodal evaluation framework for visual code generation. The framework executes each artifact and records three staged screenshots—before, during, and after a scripted interaction—then scores the evidence and source code with a checklist-guided Multimodal LLM (MLLM) referee. ArtifactsBench contains 1,825 tasks across nine domains and evaluates more than 30 LLMs. On a 280-query, six-model expert study, three staged screenshots yield 90.95% referee–human Pair ACC. Across the models shared with WebDev Arena, ArtifactsBench reaches 94.4% normalized Footrule consistency with that public human-preference ranking. We release the benchmark, evaluation harness, and baseline results at https://artifactsbenchmark.github.io/.

Large language models are increasingly deployed as interactive coding agents for end-to-end web development, yet existing benchmarks evaluate narrow slices of this workflow—typically text-conditioned generation with static correctness metrics—overlooking visual fidelity, interactivity, and codebase-level reasoning. We introduce WebCompass, a benchmark that covers the full web development lifecycle by spanning three input modalities (text, image, video) and three task types (generation, editing, repair). To match the diverse nature of these tasks, we pair checklist-guided LLM-as-a-Judge for patch-based evaluation with a novel Agent-as-a-Judge protocol that executes generated websites in a real browser and synthesizes targeted test cases. Experiments on ten models reveal a 25-point gap between the best closed- and open-source models, persistent weaknesses in visual aesthetics even among frontier systems, and divergent performance profiles across task types that challenge the notion of a single "best" web coding agent.

AI agents with advanced reasoning and tool-use capabilities have demonstrated impressive performance in web browsing for deep search. However, existing benchmarks such as BrowseComp primarily focus on textual content, overlooking the prevalence of multimodal content. To bridge this gap, we introduce MM-BrowseComp, a novel benchmark comprising 400 challenging, hand-crafted questions designed to evaluate multimodal retrieval and reasoning capabilities. Unlike prior work, MM-BrowseComp incorporates visual prompts and necessitates the extraction of key evidence from web images and videos to complete questions, rendering text-only approaches insufficient. Additionally, we provide a verified checklist for each question, enabling fine-grained analysis of multimodal dependencies and reasoning paths. Our comprehensive evaluation of 27 state-of-the-art models reveals that even leading models like GPT-5-High with tools achieve only 24.25% accuracy, highlighting the suboptimal multimodal browsing capabilities, establishing MM-BrowseComp as a rigorous new standard for the field.

Recent progress in large language models (LLMs) is largely driven by scaling training compute through either pre-training with next-token prediction (NTP) or post-training with reinforcement learning (RL). The former contributes to learning broad knowledge and skills from general data, while struggling with data inefficiency and catastrophic forgetting in continual learning settings. The latter incentivizes reasoning capabilities with strong generalization, but is constrained by limited data availability due to its reliance on human annotation. To alleviate these issues, we propose Reinforcement Learning on Pre-Training data (RLPT), which combines the advantages of learning from general data and RL. In particular, RLPT derives reward signals directly from general text data through a next-segment reasoning objective, rewarding the policy for correctly predicting next text segments conditioned on the prefix text. Experiments across multiple benchmarks and models demonstrate the effectiveness of . For example, RLPT yields substantial improvements in continual pre-training (+4.6%) and provides a strong foundation for post-training (+3.4%) on Qwen3-8B-Base.

Translating natural language mathematical statements into formal, executable code is a fundamental challenge in automated theorem proving. While prior work has focused on generation and compilation success, little attention has been paid to the critic phase—the evaluation of whether generated formalizations truly capture the semantic intent of the original problem. In this paper, we introduce CriticLean, a novel critic-guided reinforcement learning framework that elevates the role of the critic from a passive validator to an active learning component. Specifically, first, we propose the CriticLeanGPT, trained via supervised fine-tuning and reinforcement learning, to rigorously assess the semantic fidelity of Lean 4 formalizations. Then, we introduce CriticLeanBench, a benchmark designed to measure models’ ability to distinguish semantically correct from incorrect formalizations, and demonstrate that our trained CriticLeanGPT models can significantly outperform strong open- and closed-source baselines. Building on the CriticLean framework, we construct FineLeanCorpus, a dataset comprising over 509K problems that exhibits rich domain diversity, broad difficulty coverage, and high correctness based on human evaluation.Overall, our findings highlight that optimizing the critic phase is essential for producing reliable formalizations and we hope our CriticLean will provide valuable insights for future advances in formal mathematical reasoning.

While Large Language Models (LLMs) excel at algorithmic code generation, they struggle with front-end development, where correctness is judged on rendered pixels and interaction. We present ReLook, an agentic, vision-grounded reinforcement learning framework that empowers an agent to close a robust generate–diagnose–refine loop by invoking a multimodal LLM (MLLM) as a tool. During training, the agent employs an MLLM-in-the-loop to serve as a visual critic, evaluating code via screenshots and providing actionable feedback. Crucially, we enforce a strict zero-reward policy for invalid renders to guarantee renderability and mitigate reward hacking. To prevent behavioral collapse, we introduce Forced Optimization, a strict acceptance rule that admits only improving revisions, yielding monotonically better trajectories. At inference, we decouple the critic and run a lightweight, critic-free self-edit cycle, keeping latency comparable to base decoding while retaining most of the gains. Across three widely used benchmarks, ReLook consistently outperforms strong baselines in vision-grounded front-end code generation, highlighting the benefits of agentic perception, visual rewards, and training–inference decoupling.

Code LLMs have been widely used in various domains, including code generation, logical reasoning, and agent systems. However, open-access code LLMs mostly only release weights, lacking key features such as reproducible data pipelines and transparent training protocols, which are crucial for advancing deeper, more reliable investigations. To address the gap, we introduce OpenCoder, a top-tier code LLM that not only achieves performance comparable to leading models but also serves as an “open cookbook” for the research community. Unlike most prior efforts, we release not only model weights and inference code, but also the reproducible training data, complete data processing pipeline, rigorous experimental ablation results, and detailed training protocols for open scientific research. Our work identifies the key ingredients for building a top-tier code LLM: optimized heuristic rules for data cleaning and deduplication, effective recall of code-related text corpus, and high-quality synthetic data for both annealing and supervised fine-tuning stages. By offering this level of openness, we aim to broaden access to all aspects of a top-tier code LLM, with OpenCoder serving as both a powerful model and an open foundation to accelerate research and enable reproducible advancements in code intelligence. The released resource is available at https://opencoder-llm.github.io.