Correct Prediction, Wrong Steps? Consensus Reasoning Knowledge Graph for Robust Chain-of-Thought Synthesis

TL;DR

CRAFT constructs a Reasoning Knowledge Graph from multiple candidate traces to improve the quality and accuracy of LLM reasoning, outperforming existing methods on logical and mathematical benchmarks.

Contribution

Introduces CRAFT, a novel framework that mitigates reasoning flaws by synthesizing high-quality traces from consensus parts of multiple candidates.

Findings

Improves label-prediction accuracy by over 10% on average.

Outperforms all baselines across logical and mathematical reasoning benchmarks.

Enhances the quality of reasoning traces in multiple evaluation dimensions.

Abstract

LLM reasoning traces suffer from complex flaws -- *Step Internal Flaws* (logical errors, hallucinations, etc.) and *Step-wise Flaws* (overthinking, underthinking), which vary by sample. A natural approach would be to provide ground-truth labels to guide LLMs' reasoning. Contrary to intuition, we show that this yields no improvement in reasoning ability. We then propose CRAFT, a unified framework that mitigates both types of Step flaws, which builds a Reasoning Knowledge Graph (RKG) based on the consensus parts of multiple candidate traces, and synthesizes a high-quality trace through topological generation. Our approach improves label-prediction accuracy by 10+% on average, and consistently outperforms all baselines across both logical and mathematical reasoning benchmarks. Further, detailed benchmark evaluation proves that our method also improves the quality of LLMs' reasoning traces in multiple dimensions.