Reasoning Error from Known Fact: Step-Level Self-Consistency Group Relative Policy Optimization for LLMs
LLMs often produce 'context-sensitive factual hallucinations' during long reasoning by making errors due to surrounding context, even when they know the facts. This work introduces Step-level Self-Consistency Group Relative Policy Optimization (SSC-GRPO), a reinforcement learning method that assigns rewards based on how consistent individual reasoning steps are across multiple trials. This approach improves mathematical and hallucination performance.