Jun Shen

Zhiheng Xi, Xin Guo, Yang Nan et al.

Reinforcement learning (RL) has recently become the core paradigm for aligning and strengthening large language models (LLMs). Yet, applying RL in off-policy settings--where stale data from past policies are used for training--improves sample efficiency, but remains challenging: policy entropy declines sharply, optimization often becomes unstable and may even collapse. Through theoretical and empirical analysis, we identify two key insights: (i) an imbalance in optimization, where negative-advantage samples dominate the policy gradient, suppressing useful behaviors and risking gradient explosions; and (ii) the derived Entropy-Clip Rule, which reveals that the fixed clipping mechanism in PPO-like objectives systematically blocks entropy-increasing updates, thereby driving the policy toward over-exploitation at the expense of exploration. Building on these insights, we propose BAlanced Policy Optimization with Adaptive Clipping (BAPO), a simple yet effective method that dynamically adjusts clipping bounds to adaptively re-balance positive and negative contributions, preserve entropy, and stabilize RL optimization. Across diverse off-policy scenarios--including sample replay and partial rollout--BAPO achieves fast, stable, and data-efficient training. On AIME 2024 and AIME 2025 benchmarks, our 7B BAPO model surpasses open-source counterparts such as SkyWork-OR1-7B, while our 32B BAPO model not only achieves state-of-the-art results among models of the same scale but also outperforms leading proprietary systems like o3-mini and Gemini-2.5-Flash-Thinking.

Peiran Wu, Zhuorui Yu, Yunze Liu et al.

The rapid progress of large language models (LLMs) has laid the foundation for multimodal models. However, visual language models (VLMs) still face heavy computational costs when extended from images to videos due to high frame rates and long durations. Token compression is a promising solution, yet most existing training-free methods cause information loss and performance degradation. To overcome this, we propose \textbf{Memory-Augmented Reinforcement Learning-based Token Compression (MARC)}, which integrates structured retrieval and RL-based distillation. MARC adopts a \textit{retrieve-then-compress} strategy using a \textbf{Visual Memory Retriever (VMR)} to select key clips and a \textbf{Compression Group Relative Policy Optimization (C-GRPO)} framework to distil reasoning ability from a teacher to a student model. Experiments on six video benchmarks show that MARC achieves near-baseline accuracy using only one frame's tokens -- reducing visual tokens by \textbf{95\%}, GPU memory by \textbf{72\%}, and latency by \textbf{23.9\%}. This demonstrates its potential for efficient, real-time video understanding in resource-constrained settings such as video QA, surveillance, and autonomous driving.

Haonan Wang, Jiaji Mao, Lehan Wang et al.

Medical AI assistants support doctors in disease diagnosis, medical image analysis, and report generation. However, they still face significant challenges in clinical use, including limited accuracy with multimodal content and insufficient validation in real-world settings. We propose RCMed, a full-stack AI assistant that improves multimodal alignment in both input and output, enabling precise anatomical delineation, accurate localization, and reliable diagnosis through hierarchical vision-language grounding. A self-reinforcing correlation mechanism allows visual features to inform language context, while language semantics guide pixel-wise attention, forming a closed loop that refines both modalities. This correlation is enhanced by a color region description strategy, translating anatomical structures into semantically rich text to learn shape-location-text relationships across scales. Trained on 20 million image-mask-description triplets, RCMed achieves state-of-the-art precision in contextualizing irregular lesions and subtle anatomical boundaries, excelling in 165 clinical tasks across 9 modalities. It achieved a 23.5% relative improvement in cell segmentation from microscopy images over prior methods. RCMed's strong vision-language alignment enables exceptional generalization, with state-of-the-art performance in external validation across 20 clinically significant cancer types, including novel tasks. This work demonstrates how integrated multimodal models capture fine-grained patterns, enabling human-level interpretation in complex scenarios and advancing human-centric AI healthcare.