A Gradient Perspective on RLVR Stability and Winner Advantage Policy Optimization

Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but GRPO-style optimization remains prone to collapse. We analyse this instability through token-level gradient dynamics, deriving a taxonomy that predicts how updates affect next-token probabilities and entropy. The taxonomy shows that stability depends jointly on the advantage sign and token distribution under the current policy. Motivated by this finding, we propose Winner Advantage Policy Optimization (WAPO), a simple online clipped policy-gradient objective that updates only on positive-advantage completions. Across mathematical reasoning and multi-hop QA benchmarks, WAPO improves training stability and matches or outperforms baselines across multiple model families. Full code can be found at https://github.com/layer6ai-labs/wapo.

The Uncertainty-based Retrieval Framework for Ancient Chinese CWS and POS

Automatic analysis for modern Chinese has greatly improved the accuracy of text mining in related fields, but the study of ancient Chinese is still relatively rare. Ancient text division and lexical annotation are important parts of classical literature comprehension, and previous studies have tried to construct auxiliary dictionary and other fused knowledge to improve the performance. In this paper, we propose a framework for ancient Chinese Word Segmentation and Part-of-Speech Tagging that makes a twofold effort: on the one hand, we try to capture the wordhood semantics; on the other hand, we re-predict the uncertain samples of baseline model by introducing external knowledge. The performance of our architecture outperforms pre-trained BERT with CRF and existing tools such as Jiayan.