Navigating the complexities of physics reasoning has long been a difficult task for Large Language Models (LLMs), requiring a synthesis of profound conceptual understanding and adept problem-solving techniques. In this study, we investigate the application of advanced instruction-tuned reasoning models, such as Deepseek-R1, to address a diverse spectrum of physics problems curated from the challenging SciBench benchmark. Our comprehensive experimental evaluation reveals the remarkable capabilities of reasoning models. Not only do they achieve state-of-the-art accuracy in answering intricate physics questions, but they also generate distinctive reasoning patterns that emphasize on symbolic derivation. Furthermore, our findings indicate that even for these highly sophisticated reasoning models, the strategic incorporation of few-shot prompting can still yield measurable improvements in overall accuracy, highlighting the potential for continued performance gains.

Transcending human cognitive limitations represents a critical frontier in LLM training. Proprietary agentic systems like DeepResearch have demonstrated superhuman capabilities on extremely complex information-seeking benchmarks such as BrowseComp, a feat previously unattainable. We posit that their success hinges on a sophisticated reasoning pattern absent in open-source models: the ability to systematically reduce extreme uncertainty when navigating vast information landscapes. Based on this insight, we introduce WebSailor, a complete post-training methodology designed to instill this crucial capability. Our approach involves generating novel, high-uncertainty tasks through structured sampling and information obfuscation, RFT cold start, and an efficient agentic RL training algorithm, Duplicating Sampling Policy Optimization (DUPO). With this integrated pipeline, WebSailor significantly outperforms all opensource agents in complex information-seeking tasks, matching proprietary agents' performance and closing the capability gap.

In this work, we investigate whether improving task clarity can enhance reasoning ability of large language models, focusing on theorem proving in Coq. We introduce a concept-level metric to evaluate task clarity and show that adding structured semantic context to the standard input used by modern LLMs, leads to a 1.85$\times$ improvement in clarity score (44.5\%~$\rightarrow$~82.3\%). Using the general-purpose model \texttt{DeepSeek-V3}, our approach leads to a 2.1$\times$ improvement in proof success (21.8\%~$\rightarrow$~45.8\%) and outperforms the previous state-of-the-art \texttt{Graph2Tac} (33.2\%). We evaluate this on 1,386 theorems randomly sampled from 15 standard Coq packages, following the same evaluation protocol as \texttt{Graph2Tac}. Furthermore, fine-tuning smaller models on our structured data can achieve even higher performance (48.6\%). Our method uses selective concept unfolding to enrich task descriptions, and employs a Planner--Executor architecture. These findings highlight the value of structured task representations in bridging the gap between understanding and reasoning.

Reasoning Language Models (RLMs) have gained traction for their ability to perform complex, multi-step reasoning tasks through mechanisms such as Chain-of-Thought (CoT) prompting or fine-tuned reasoning traces. While these capabilities promise improved reliability, their impact on robustness to social biases remains unclear. In this work, we leverage the CLEAR-Bias benchmark, originally designed for Large Language Models (LLMs), to investigate the adversarial robustness of RLMs to bias elicitation. We systematically evaluate state-of-the-art RLMs across diverse sociocultural dimensions, using an LLM-as-a-judge approach for automated safety scoring and leveraging jailbreak techniques to assess the strength of built-in safety mechanisms. Our evaluation addresses three key questions: (i) how the introduction of reasoning capabilities affects model fairness and robustness; (ii) whether models fine-tuned for reasoning exhibit greater safety than those relying on CoT prompting at inference time; and (iii) how the success rate of jailbreak attacks targeting bias elicitation varies with the reasoning mechanisms employed. Our findings reveal a nuanced relationship between reasoning capabilities and bias safety. Surprisingly, models with explicit reasoning, whether via CoT prompting or fine-tuned reasoning traces, are generally more vulnerable to bias elicitation than base models without such mechanisms, suggesting reasoning may unintentionally open new pathways for stereotype reinforcement. Reasoning-enabled models appear somewhat safer than those relying on CoT prompting, which are particularly prone to contextual reframing attacks through storytelling prompts, fictional personas, or reward-shaped instructions. These results challenge the assumption that reasoning inherently improves robustness and underscore the need for more bias-aware approaches to reasoning design.

Reinforcement learning with verifiable rewards (RL VR) is a promising approach for improving the complex reasoning abilities of large language models (LLMs). However, current RL VR methods face two significant challenges: the near-miss reward problem, where a small mistake can invalidate an otherwise correct reasoning process, greatly hindering training efficiency; and exploration stagnation, where models tend to focus on solutions within their "comfort zone," lacking the motivation to explore potentially more effective alternatives. To address these challenges, we propose StepHint, a novel RL VR algorithm that utilizes multi-level stepwise hints to help models explore the solution space more effectively. StepHint generates valid reasoning chains from stronger models and partitions these chains into reasoning steps using our proposed adaptive partitioning method. The initial few steps are used as hints, and simultaneously, multiple-level hints (each comprising a different number of steps) are provided to the model. This approach directs the model's exploration toward a promising solution subspace while preserving its flexibility for independent exploration. Additionally, the external reasoning pathways help the model develop better reasoning abilities, enabling it to move beyond its "comfort zone" and mitigate exploration stagnation. StepHint outperforms competitive RL VR enhancement methods across six mathematical benchmarks, while also demonstrating superior generalization and excelling over baselines on out-of-domain benchmarks. Eliciting the reasoning capabilities of large language models (LLMs) through Reinforcement Learning with V er-ifiable Rewards (RL VR) has emerged as a powerful paradigm (Jaech et al., 2024; Guo et al., 2025a). In RL VR frameworks, a policy model explores the solution space by generating reasoning chains.

Information retrieval is a cornerstone of modern knowledge acquisition, enabling billions of queries each day across diverse domains. However, traditional keyword-based search engines are increasingly inadequate for handling complex, multi-step information needs. Our position is that Large Language Models (LLMs), endowed with reasoning and agentic capabilities, are ushering in a new paradigm termed Agentic Deep Research. These systems transcend conventional information search techniques by tightly integrating autonomous reasoning, iterative retrieval, and information synthesis into a dynamic feedback loop. We trace the evolution from static web search to interactive, agent-based systems that plan, explore, and learn. We also introduce a test-time scaling law to formalize the impact of computational depth on reasoning and search. Supported by benchmark results and the rise of open-source implementations, we demonstrate that Agentic Deep Research not only significantly outperforms existing approaches, but is also poised to become the dominant paradigm for future information seeking. All the related resources, including industry products, research papers, benchmark datasets, and open-source implementations, are collected for the community in https://github.com/DavidZWZ/Awesome-Deep-Research.

While recent advances in preference learning have enhanced alignment in human feedback, mathematical reasoning remains a persistent challenge. We investigate how data diversification strategies in preference optimization can improve the mathematical reasoning abilities of large language models (LLMs). We evaluate three common data generation methods: temperature sampling, Chain-of-Thought prompting, and Monte Carlo Tree Search (MCTS), and introduce Diversified-ThinkSolve (DTS), a novel structured approach that systematically decomposes problems into diverse reasoning paths. Our results show that with strategically diversified preference data, models can substantially improve mathematical reasoning performance, with the best approach yielding gains of 7.1% on GSM8K and 4.2% on MATH over the base model. Despite its strong performance, DTS incurs only a marginal computational overhead (1.03x) compared to the baseline, while MCTS is nearly five times more costly with lower returns. These findings demonstrate that structured exploration of diverse problem-solving methods creates more effective preference data for mathematical alignment than traditional approaches.

Recent advances in Multi-Modal Large Language Models (MLLMs) have enabled unified processing of language, vision, and structured inputs, opening the door to complex tasks such as logical deduction, spatial reasoning, and scientific analysis. Despite their promise, the reasoning capabilities of MLLMs, particularly those augmented with intermediate thinking traces (MLLMs-T), remain poorly understood and lack standardized evaluation benchmarks. Existing work focuses primarily on perception or final answer correctness, offering limited insight into how models reason or fail across modalities. To address this gap, we introduce the MMLU-Reason, a new benchmark designed to rigorously evaluate multi-modal reasoning with explicit thinking. The MMLU-Reason comprises 1) a high-difficulty dataset of 1,083 questions spanning six diverse reasoning types with symbolic depth and multi-hop demands and 2) a modular Reasoning Trace Evaluation Pipeline (RTEP) for assessing reasoning quality beyond accuracy through metrics like relevance, consistency, and structured error annotations. Empirical results show that MLLMs-T overall outperform non-thinking counterparts, but even top models like Claude-3.7-Sonnet and Gemini-2.5 Pro suffer from reasoning pathologies such as inconsistency and overthinking. This benchmark reveals persistent gaps between accuracy and reasoning quality and provides an actionable evaluation pipeline for future model development. Overall, the MMLU-Reason offers a scalable foundation for evaluating, comparing, and improving the next generation of multi-modal reasoning systems.

Large Language Models (LLMs) have emerged as one of the most significant technological advancements in artificial intelligence in recent years. Their ability to understand, generate, and reason with natural language has transformed how we interact with AI systems. With the development of LLM-based agents and reinforcement-learning-based reasoning models, the study of applying reinforcement learning in agent frameworks has become a new research focus. However, all previous studies face the challenge of deciding the tool calling process and the reasoning process simultaneously, and the chain of reasoning was solely relied on the unprocessed raw result with redundant information and symbols unrelated to the task from the tool, which impose a heavy burden on the model's capability to reason. Therefore, in our research, we proposed a hierarchical framework Agent-as-tool that detach the tool calling process and the reasoning process, which enables the model to focus on the verbally reasoning process while the tool calling process is handled by another agent. Our work had achieved comparable results with only a slight reinforcement fine-tuning on 180 samples, and had achieved exceptionally well performance in Bamboogle with 63.2% of exact match and 75.2% in cover exact match, exceeding Search-R1 by 4.8% in exact match and 3.2% in cover exact match.

With the development of generative artificial intelligence and instruction tuning techniques, multimodal large language models (MLLMs) have made impressive progress on general reasoning tasks. Benefiting from the chain-of-thought (CoT) methodology, MLLMs can solve the visual reasoning problem step-by-step. However, existing MLLMs still face significant challenges when applied to pathology visual reasoning tasks: (1) LLMs often underperforms because they lack domain-specific information, which can lead to model hallucinations. (2) The additional reasoning steps in CoT may introduce errors, leading to the divergence of answers. To address these limitations, we propose PathCoT, a novel zero-shot CoT prompting method which integrates the pathology expert-knowledge into the reasoning process of MLLMs and incorporates self-evaluation to mitigate divergence of answers. Specifically, PathCoT guides the MLLM with prior knowledge to perform as pathology experts, and provides comprehensive analysis of the image with their domain-specific knowledge. By incorporating the experts' knowledge, PathCoT can obtain the answers with CoT reasoning. Furthermore, PathCoT incorporates a self-evaluation step that assesses both the results generated directly by MLLMs and those derived through CoT, finally determining the reliable answer. The experimental results on the PathMMU dataset demonstrate the effectiveness of our method on pathology visual understanding and reasoning.