NCV: A Node-Wise Consistency Verification Approach for Low-Cost Structured Error Localization in LLM Reasoning

2025-10-03

AI Summary

To address inaccurate error localization and high verification costs in multi-step reasoning of large language models (LLMs), this paper proposes the Node-level Consistency Verification (NCV) framework. NCV decomposes reasoning chains into fine-grained, verifiable nodes and formulates verification as a zero-shot, training-free, lightweight binary classification task—assessing consistency at each node to enable precise error localization. Unlike conventional chain-level evaluation or costly multiple-sampling approaches, NCV eliminates long-sequence generation and mitigates attention dilution. On public benchmarks, NCV improves F1 scores by 10–25% over baselines while reducing token consumption to only 1/6–1/58 of that required by standard methods. This yields substantial gains in verification efficiency, accuracy, and interpretability, offering a scalable and principled solution for LLM reasoning validation.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Verifying multi-step reasoning in large language models is difficult due to imprecise error localization and high token costs. Existing methods either assess entire reasoning chains, suffering attention dilution, or rely on expensive multi-sampling. We introduce Node-wise Consistency Verification (NCV), a training-free framework that recasts verification as lightweight binary consistency checks at the node level. By decomposing the chain of thought into interconnected verification nodes, NCV precisely localizes errors and avoids unnecessary long-form generation. Experiments demonstrate that our approach enhances interpretability and efficiency, presenting a scalable solution for reliable LLM reasoning verification. On public datasets, NCV achieves a 10% to 25% improvement in F1 scores over baselines while utilizing $6 imes$~$58 imes$ fewer tokens than traditional methods like CoT-based verifiers.

Problem

• Verifying multi-step reasoning in large language models
• Imprecise error localization with high token costs
• Assessing reasoning chains suffers from attention dilution

Innovation

• Node-wise binary consistency checks for verification
• Training-free framework avoiding long-form generation
• Lightweight decomposition of reasoning chains into nodes