Can I read Correctness-Aware Advantage Shaping for LLMs on EtoBox?
Correctness-Aware Advantage Shaping for LLMs by yangding2022 is a document available to read on EtoBox.
What is Correctness-Aware Advantage Shaping for LLMs about?
The document introduces Low-Entropy Segment Shaping (LESS), a reinforcement learning framework designed to enhance the reasoning abilities of large language models by focusing on low-entropy segments that contribute to correct responses. LESS modifies token-level advantages based on the correctness of segments, amplifying those unique to correct answers while suppressing those linked to incorrect ones. Experiments demonstrate that LESS consistently improves accuracy across various mathematical reasoning ben
- Author
- yangding2022
- Language
- EN