Rethinking Token-Level Credit Assignment in RLVR: A Polarity-Entropy Analysis

Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved the reasoning ability of Large Language Models (LLMs). However, its sparse outcome-based rewards pose a fundamental credit assignment problem. We analyze this problem through the joint lens of reward polarity and token entropy. Our diagnostic tool, the Four Quadrant Decomposition, isolates token updates by polarity and entropy, and controlled ablations show that reasoning improvements concentrate in the high-entropy quadrants. To justify this observation theoretically, we adapt Conditional Mutual Information to the autoregressive RLVR setting and prove that the credit a token can carry is upper-bounded by its entropy. This view yields testable predictions that reasoning gains arise primarily from high-entropy tokens, with unique roles for positive and negative updates. A gradient analysis of GRPO further reveals how uniform reward broadcast dilutes signal at high-entropy positions while over-crediting deterministic tokens. Grounded in these insights, we propose Entropy-Aware Policy Optimization (EAPO) that modulates token-level learning signals accordingly. Extensive experiments demonstrate that EAPO outperforms strong baselines across two model families.

GLM-5: from Vibe Coding to Agentic Engineering

We present GLM-5, a next-generation foundation model designed to transition the paradigm of vibe coding to agentic engineering. Building upon the agentic, reasoning, and coding (ARC) capabilities of its predecessor, GLM-5 adopts DSA to significantly reduce training and inference costs while maintaining long-context fidelity. To advance model alignment and autonomy, we implement a new asynchronous reinforcement learning infrastructure that drastically improves post-training efficiency by decoupling generation from training. Furthermore, we propose novel asynchronous agent RL algorithms that further improve RL quality, enabling the model to learn from complex, long-horizon interactions more effectively. Through these innovations, GLM-5 achieves state-of-the-art performance on major open benchmarks. Most critically, GLM-5 demonstrates unprecedented capability in real-world coding tasks, surpassing previous baselines in handling end-to-end software engineering challenges. Code, models, and more information are available at https://github.com/zai-org/GLM-5.

ClinKD: Cross-Modal Clinic Knowledge Distiller For Multi-Task Medical Images

Med-VQA (Medical Visual Question Answering) is a crucial subtask within the broader VQA (Visual Question Answering) domain. This task requires a visual question answering system to analyze the provided image and corresponding question,offering reasonable analysis and suggestions to assist medical professionals in making pathological diagnoses, or ideally, enabling the system to independently provide correct diagnoses. Furthermore, more advanced Med-VQA tasks involve Referring and Grounding, which not only require the system to accurately comprehend medical images but also to pinpoint specific biological locations within those images. While many large pre-trained models have demonstrated substantial VQA capabilities,challenges persist in the medical imaging domain. The intricacy of biological features in medical images and the scarcity of high-quality medical image datasets, combined with the fact that current models are not tailored for the medical field in terms of architecture and training paradigms, hinder the full exploitation of model generalization. This results in issues such as hallucination in Visual Grounding. In this paper, we introduce the ClinKD model, which incorporates modifications to model position encoding and a diversified training process. Initially, we enhance the model's ability to perceive image and modality variations by using Med-CLIP Guided Rotary Position Embedding. Subsequently, we leverage distillation to provide prior knowledge to the model before using complete training data. Additionally, the feedback-based training process during the formal training phase further enhances data utilization. Notably, under unchanged evaluation protocols, we achieve a new state-of-the-art performance on the Med-GRIT-270k dataset, and the Med-CLIP Guided Rotary Position Embedding approach presents potential for generalizing to universal model position encoding.