New method uses hidden states to improve AI reasoning credit assignment

Researchers have developed a new method called Span-level Hidden state Enabled Advantage Reweighting (SHEAR) to improve credit assignment in reinforcement learning for language models. SHEAR leverages the Wasserstein distance between hidden state distributions of correct and incorrect reasoning paths to identify and amplify learning signals in critical token areas. This approach requires no additional annotation or reward model training, demonstrating improved performance on mathematical reasoning and code generation tasks compared to existing methods. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a novel, annotation-free method to improve AI reasoning and code generation capabilities.

RANK_REASON Academic paper introducing a novel method for improving AI training.

Read on arXiv cs.CL →

• SHEAR
• Wasserstein distance
• reinforcement learning
• language models
• arXiv
• GRPO
• RLVR

• paper
• safety