Yue Yang

Shenzhi Wang, Le Yu, Chang Gao et al. · tsinghua

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful approach to enhancing the reasoning capabilities of Large Language Models (LLMs), while its mechanisms are not yet well understood. In this work, we undertake a pioneering exploration of RLVR through the novel perspective of token entropy patterns, comprehensively analyzing how different tokens influence reasoning performance. By examining token entropy patterns in Chain-of-Thought (CoT) reasoning, we observe that only a small fraction of tokens exhibit high entropy, and these tokens act as critical forks that steer the model toward diverse reasoning pathways. Furthermore, studying how entropy patterns evolve during RLVR training reveals that RLVR largely adheres to the base model's entropy patterns, primarily adjusting the entropy of high-entropy tokens. These findings highlight the significance of high-entropy tokens (i.e., forking tokens) to RLVR. We ultimately improve RLVR by restricting policy gradient updates to forking tokens and uncover a finding even beyond the 80/20 rule: utilizing only 20% of the tokens while maintaining performance comparable to full-gradient updates on the Qwen3-8B base model and significantly surpassing full-gradient updates on the Qwen3-32B (+11.04 on AIME'25 and +7.71 on AIME'24) and Qwen3-14B (+4.79 on AIME'25 and +5.21 on AIME'24) base models, highlighting a strong scaling trend. In contrast, training exclusively on the 80% lowest-entropy tokens leads to a marked decline in performance. These findings indicate that the efficacy of RLVR primarily arises from optimizing the high-entropy tokens that decide reasoning directions. Collectively, our results highlight the potential to understand RLVR through a token-entropy perspective and optimize RLVR by leveraging high-entropy minority tokens to further improve LLM reasoning.

Xuehai He, Shu Chen, Zeqian Ju et al.

Medical dialogue systems are promising in assisting in telemedicine to increase access to healthcare services, improve the quality of patient care, and reduce medical costs. To facilitate the research and development of medical dialogue systems, we build two large-scale medical dialogue datasets: MedDialog-EN and MedDialog-CN. MedDialog-EN is an English dataset containing 0.3 million conversations between patients and doctors and 0.5 million utterances. MedDialog-CN is an Chinese dataset containing 1.1 million conversations and 4 million utterances. To our best knowledge, MedDialog-(EN,CN) are the largest medical dialogue datasets to date. The dataset is available at https://github.com/UCSD-AI4H/Medical-Dialogue-System

Andrew Zhao, Yiran Wu, Yang Yue et al. · tsinghua

Reinforcement learning with verifiable rewards (RLVR) has shown promise in enhancing the reasoning capabilities of large language models by learning directly from outcome-based rewards. Recent RLVR works that operate under the zero setting avoid supervision in labeling the reasoning process, but still depend on manually curated collections of questions and answers for training. The scarcity of high-quality, human-produced examples raises concerns about the long-term scalability of relying on human supervision, a challenge already evident in the domain of language model pretraining. Furthermore, in a hypothetical future where AI surpasses human intelligence, tasks provided by humans may offer limited learning potential for a superintelligent system. To address these concerns, we propose a new RLVR paradigm called Absolute Zero, in which a single model learns to propose tasks that maximize its own learning progress and improves reasoning by solving them, without relying on any external data. Under this paradigm, we introduce the Absolute Zero Reasoner (AZR), a system that self-evolves its training curriculum and reasoning ability by using a code executor to both validate proposed code reasoning tasks and verify answers, serving as an unified source of verifiable reward to guide open-ended yet grounded learning. Despite being trained entirely without external data, AZR achieves overall SOTA performance on coding and mathematical reasoning tasks, outperforming existing zero-setting models that rely on tens of thousands of in-domain human-curated examples. Furthermore, we demonstrate that AZR can be effectively applied across different model scales and is compatible with various model classes.

Matthieu Lin, Jenny Sheng, Andrew Zhao et al. · tsinghua

This survey organizes the intricate literature on the design and optimization of emerging structures around post-trained LMs. We refer to this overarching structure as scaffolded LMs and focus on LMs that are integrated into multi-step processes with tools. We view scaffolded LMs as semi-parametric models wherein we train non-parametric variables, including the prompt, tools, and scaffold's code. In particular, they interpret instructions, use tools, and receive feedback all in language. Recent works use an LM as an optimizer to interpret language supervision and update non-parametric variables according to intricate objectives. In this survey, we refer to this paradigm as training of scaffolded LMs with language supervision. A key feature of non-parametric training is the ability to learn from language. Parametric training excels in learning from demonstration (supervised learning), exploration (reinforcement learning), or observations (unsupervised learning), using well-defined loss functions. Language-based optimization enables rich, interpretable, and expressive objectives, while mitigating issues like catastrophic forgetting and supporting compatibility with closed-source models. Furthermore, agents are increasingly deployed as co-workers in real-world applications such as Copilot in Office tools or software development. In these mixed-autonomy settings, where control and decision-making are shared between human and AI, users point out errors or suggest corrections. Accordingly, we discuss agents that continuously improve by learning from this real-time, language-based feedback and refer to this setting as streaming learning from language supervision.

Nguyen Van Hoang, Soeren Hougaard Mulvad, Dexter Neo Yuan Rong et al.

A significant shortcoming of current state-of-the-art (SOTA) named-entity recognition (NER) systems is their lack of generalization to unseen domains, which poses a major problem since obtaining labeled data for NER in a new domain is expensive and time-consuming. We propose ZERO, a model that performs zero-shot and few-shot learning in NER to generalize to unseen domains by incorporating pre-existing knowledge in the form of semantic word embeddings. ZERO first obtains contextualized word representations of input sentences using the model LUKE, reduces their dimensionality, and compares them directly with the embeddings of the external knowledge, allowing ZERO to be trained to recognize unseen output entities. We find that ZERO performs well on unseen NER domains with an average macro F1 score of 0.23, outperforms LUKE in few-shot learning, and even achieves competitive scores on an in-domain comparison. The performance across source-target domain pairs is shown to be inversely correlated with the pairs' KL divergence.

Anli Lim, Bharath Ramesh, Yue Yang et al.

This paper describes the development of a novel algorithm to tackle the problem of real-time video stabilization for unmanned aerial vehicles (UAVs). There are two main components in the algorithm: (1) By designing a suitable model for the global motion of UAV, the proposed algorithm avoids the necessity of estimating the most general motion model, projective transformation, and considers simpler motion models, such as rigid transformation and similarity transformation. (2) To achieve a high processing speed, optical-flow based tracking is employed in lieu of conventional tracking and matching methods used by state-of-the-art algorithms. These two new ideas resulted in a real-time stabilization algorithm, developed over two phases. Stage I considers processing the whole sequence of frames in the video while achieving an average processing speed of 50fps on several publicly available benchmark videos. Next, Stage II undertakes the task of real-time video stabilization using a multi-threading implementation of the algorithm designed in Stage I.