Reproduce “CER: Confidence Enhanced Reasoning in LLMs” on OpenResearch

About this paper

Ali Razghandi, Seyed Mohammad Hadi Hosseini, Mahdieh Soleymani Baghshah

The Confidence Enhanced Reasoning (CER) framework enhances Large Language Model performance in complex reasoning by estimating confidence from critical tokens in intermediate steps, then aggregating these to weight and select final answers. This approach yielded average accuracy increases of up to 5.3% on mathematical reasoning and 3.0% on knowledge-intensive question answering datasets across various models.

Key insights

• Integrating LLM confidence at critical decision points within a multi-step reasoning process significantly improves the accuracy and reliability of responses.
• Focusing confidence estimation on specific, identifiable "critical tokens" within intermediate steps provides a more effective and less noisy probabilistic signal than attempting to evaluate confidence for every token.
• Incorporating confidence signals from all intermediate reasoning steps, rather than just the final answer, is crucial for achieving optimal performance gains.

Summary by alphaXiv.