Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Ling, Zipeng, Liu, Shuliang, Fu, Shenghong, Tang, Yuehao, Son, Seonil, Wan, Yao, Hu, Xuming
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2604.14121
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866913035781668864
author Ling, Zipeng
Liu, Shuliang
Fu, Shenghong
Tang, Yuehao
Son, Seonil
Wan, Yao
Hu, Xuming
author_facet Ling, Zipeng
Liu, Shuliang
Fu, Shenghong
Tang, Yuehao
Son, Seonil
Wan, Yao
Hu, Xuming
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2604_14121
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Correct Prediction, Wrong Steps? Consensus Reasoning Knowledge Graph for Robust Chain-of-Thought Synthesis
Ling, Zipeng
Liu, Shuliang
Fu, Shenghong
Tang, Yuehao
Son, Seonil
Wan, Yao
Hu, Xuming
Computation and Language
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.
title Correct Prediction, Wrong Steps? Consensus Reasoning Knowledge Graph for Robust Chain-of-Thought Synthesis
topic Computation and Language
url https://arxiv.org/abs/2604.14121