Researchers have introduced the Relative Surprisal Index (RSI), a new metric for Reinforcement Learning with Verifiable Rewards (RLVR) in large language models. RSI aims to reconcile conflicting approaches in RLVR by considering both token entropy and probability. The proposed RSI Selection (RSI-S) method filters tokens within a stable RSI interval, removing both redundant and unstable tokens. Empirical results show RSI-S improves accuracy on benchmarks like AIME and AMC across various Qwen2.5 model scales. AI
IMPACT Introduces a novel metric and filtering method that could improve LLM reasoning capabilities in RLVR applications.
RANK_REASON The cluster contains an academic paper detailing a new metric and method for improving LLM reasoning.
- AMC
- GRPO
- large-language models
- Qwen2.5-1.5B
- Qwen2.5-3B
- qwen2.5:7b
- Relative Surprisal Index
- RLVR
- RSI Selection
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
