Firehose

30 JUL 2026 · Paper

Not All Tokens Deserve Equal Credit: Counterfactual Sensitivity Credit Reallocation for Long-CoT Reasoning

This paper proposes a method to improve long-CoT reasoning in large language models by addressing the issue of unequal token contributions to the final outcome. It shows that current methods, such as GRPO, assign too much credit to highly sensitive tokens and proposes a new method, CSCR, that reduces credit for these tokens to improve performance.

reinforcement learning verifiable rewards long-CoT reasoning continual learning counterfactual sensitivity token-level learning value

Browse by tag

Not All Tokens Deserve Equal Credit: Counterfactual Sensitivity Credit Reallocation for Long-CoT Reasoning