Agentic reinforcement learning enables large language models to solve long-horizon tasks by interacting with the environment and internalizing tool-use behavior into their reasoning. Prior work assigns supervision primarily based on outcome rewards or external reward models, but largely ignores environment observations, a critical source of learning. Consequently, agents may identify successful actions without understanding how the environment responds, producing suboptimal policies. To address this, we propose SOAR (Supervision from Observation for Agentic Reinforcement Learning), which assigns positive advantages to observation tokens proportional to the negative entropy of preceding actions. This encourages the agent to learn from outcomes of confident actions, grounding policy updates in environment dynamics and improving anticipation of tool-call consequences. Empirical results across three domains and 14 benchmarks show that SOAR improves performance, yielding gains of up to 7.0% on general reasoning tasks and 16.9% on deep research tasks, while reducing erroneous and inefficient tool usage.

Chain-of-Thought (CoT) prompting has achieved remarkable success in unlocking the reasoning capabilities of Large Language Models (LLMs). Although CoT prompting enhances reasoning, its verbosity imposes substantial computational overhead. Recent works often focus exclusively on outcome alignment and lack supervision on the intermediate reasoning process. These deficiencies obscure the analyzability of the latent reasoning chain. To address these challenges, we introduce **Render-of-Thought (RoT)**, the first framework to reify the reasoning chain by rendering textual steps into images, making the latent rationale explicit and traceable. Specifically, we leverage the vision encoders of existing Vision Language Models (VLMs) as semantic anchors to align the vision embeddings with the textual space. This design ensures **plug-and-play** implementation without incurring additional pre-training overhead. Extensive experiments on mathematical and logical reasoning benchmarks demonstrate that our method achieves 3-4× token compression and substantial inference acceleration compared to explicit CoT. Furthermore, it demonstrates a competitive efficiency-accuracy Pareto exploration compared to other methods, validating the feasibility of this paradigm. Our code is available at https://github.com/TencentBAC/RoT

Medical large Vision-Language Models (Med-LVLMs) have shown strong potential in multimodal clinical applications such as medical visual question answering and report generation. However, Med-LVLMs remain challenged by hallucinations caused by modality misalignment, where models prioritize textual knowledge over visual evidence and generate outputs that conflict with medical images. To mitigate this issue, recent studies have explored preference optimization to improve image–text alignment, achieving promising results. Despite these advances, existing preference-based methods still face two limitations in medical settings: (1) overfitting to superficial cues, and (2) pseudo convergence of the preference signal. In this paper, we propose Dynamic Evidence-Guided Preference Optimization (DEPO), a new framework that enables evidence-aware and adaptive preference learning for Med-LVLMs. DEPO introduces Multi-Modal Evidence Perturbation (MEP) to suppress non-causal textual and visual shortcuts, and Dispreferred Evidence Resampling (DER) to continuously update dispreferred responses as hallucination patterns evolve. Experiments on multiple medical VQA and report generation benchmarks demonstrate consistent improvements over existing methods, with strong robustness across datasets and architectures. All Codes and data will be released after review.

The emergence of reasoning models, exemplified by OpenAI o1, signifies a transition from intuitive to deliberative cognition, effectively reorienting the scaling laws from pre-training paradigms toward test-time computation. While Monte Carlo Tree Search (MCTS) has shown promise in this domain, existing approaches typically treat each rollout as an isolated trajectory. This lack of information sharing leads to severe inefficiency and substantial computational redundancy, as the search process fails to leverage insights from prior explorations. To address these limitations, we propose PRISM-MCTS, a novel reasoning framework that draws inspiration from human parallel thinking and reflective processes. PRISM-MCTS integrates a Process Reward Model (PRM) with a dynamic shared memory, capturing both "Heuristics" and "Fallacies". By reinforcing successful strategies and pruning error-prone branches, PRISM-MCTS effectively achieves refinement. Furthermore, we develop a data-efficient training strategy for the PRM, achieving high-fidelity evaluation under a few-shot regime. Empirical evaluations across diverse reasoning benchmarks substantiate the efficacy of PRISM-MCTS. Notably, it halves the trajectory requirements on GPQA while surpassing MCTS-RAG and Search-o1, demonstrating that it scales inference by reasoning judiciously rather than exhaustively.

SQL languages often feature nested structures that require robust interaction with databases. Aside from the well-validated schema linking methods on PLMs and LLMs, we introduce the Linearly Incremental SQL Translator (LIST), a novel algorithmic toolkit designed to leverage the notable reasoning and tool interaction capabilities inherent in LLMs. LIST transforms complex SQL queries into grammatically verifiable sub-queries which are arranged sequentially to reflect single-hop reasoning steps, enhancing both the granularity and accuracy of database interactions. With in-context learning, our experiments demonstrated significant improvements, achieving notable performance of 60.56% and 56.32% on the BIRD dataset with GPT-4o and Llama-3-70B-Instruct. To the best of our knowledge, this achieves SOTA performance among non-schema linking methods, also surpassing a series of schema linking based approaches at a comparable or better cost.

Direct Preference Optimization (DPO) has recently emerged as an efficient and effective method for aligning large language models with human preferences. However, constructing high-quality preference datasets remains challenging, often necessitating expensive manual or powerful LM annotations. Additionally, standard DPO exhibits suboptimal performance in complex reasoning tasks, such as mathematical and code reasoning. In this paper, we introduce an approach to collect preference pairs through iterative sampling and execution feedback, tailored to the current learning state (e.g. well-learned, mis-learned, and unlearned) of the policy model. To alleviate the failures of DPO and improve its applicability in reasoning tasks, we propose , an iterative uncertainty-aware preference optimization method that achieves fine-grained preference control by assessing model confidence. We validate our approach across three reasoning tasks, incorporating five established reasoning datasets and one self-curated dataset. Our experimental results demonstrate an overall improvement of 3.6% over the standard DPO method and show the model exhibits promising generalizability.

As the scale of Large Language Models (LLMs) increases, it is necessary to compress the models to reduce the substantial demand on computational resources. Network pruning significantly reduces the model size by converting the weight matrix from dense to sparse data format. Current methodologies advocate for one-shot pruning to avoid the expense of retraining, ensuring the maintenance of model performance under conditions of 50%-60% unstructured pruning. Nevertheless, matrices characterized by this level of sparsity could not be treated as sparse matrices, because the indices would incur significant costs. To mitigate this problem, NVIDIA introduced the 2:4 structured sparsity. However, we observe a notable decline in model performance when adopting 2:4 structured sparsity due to group constraints. In this paper, we introduce the Weight Recover Prune (WRP) approach. By recovering a minimal set of critical weights, WRP aims to enhance model performance while maintaining the efficiency of the compression. Our evaluation of the WRP method on the LLAMA2 and OPT models shows that it outperforms other 2:4 pattern one-shot pruning methods. Meanwhile, WRP can guarantee that the size of the pruned model is about 60% of the dense model. Our code is available at: https://github.com/TanZhendong/WRP.