The authors propose CheckRLM, a framework that enhances the reliability of Reasoning Language Models by using Retrieval-Augmented Generation to timely check and correct factual errors during inference. This approach extracts factual claims from reasoning chains to identify inconsistencies and applies minimal-cost corrections via external knowledge.
- Extracts factual claims to localize subtle knowledge inconsistencies.
- Performs precise corrections leveraging external knowledge upon error detection.
- Mitigates error accumulation in long-horizon reasoning with lower costs.
- Substantially outperforms existing baselines in extensive experiments.
CheckRLM ensures coherence between the reasoning chain and correct knowledge, addressing the issue of factual errors in knowledge-intensive tasks.