Correct Prediction, Wrong Steps? Consensus Reasoning Knowledge Graph for Robust Chain-of-Thought Synthesis
For researchers working on improving LLM reasoning, this work offers a novel method to mitigate step-level flaws without relying on ground-truth labels, achieving significant accuracy gains.
The paper identifies that providing ground-truth labels does not improve LLM reasoning, and proposes CRAFT, a framework that builds a Reasoning Knowledge Graph from consensus parts of multiple candidate traces to synthesize high-quality reasoning traces. CRAFT improves label-prediction accuracy by over 10% on average and outperforms baselines on logical and mathematical reasoning benchmarks.
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.