RG-CoT: A Rule-Guided Chain-of-Thought Evaluation Framework for Evaluating and Aligning LLM Reasoning

ACL ARR 2026 January Submission8694 Authors

Keywords: LLMs Evaluation, Chain-of-Thought, Alignment

Abstract:

While Large Language Models (LLMs) exhibit remarkable generative capabilities, evaluating their internal reasoning processes remains significantly more challenging than assessing superficial outputs. Prevailing automated metrics predominantly focus on surface-level textual quality, failing to capture critical dimensions such as logical coherence and interpretability.
To bridge this gap, we propose RG-CoT, a novel evaluation framework that employs rule-guided distillation to extract interpretable Chain-of-Thought (CoT) rationales from target models. RG-CoT establishes a multi-dimensional verification system, assessing correctness, logical consistency, reliability, format compliance, and self reflection.This enables efficient model evaluation and alignment through parameter-efficient fine-tuning on high-quality distilled data.
We conduct extensive experiments on the preprocessed COIG-PC dataset using a suite of state-of-the-art models.
Our study spans 10 task categories, involving over 180,000 distilled CoT samples and 5,500 test queries. Results demonstrate that RG-CoT effectively evaluates high-quality model data; notably, fine-tuning on merely 5.6\% of the curated high-quality data significantly outperforms training on the entire 180k corpus. Furthermore, out-of-distribution tests on the BANK-AUDIT benchmark confirm the robust generalizability of our framework.

Paper Type: Long

Research Area: Safety and Alignment in LLMs

Research Area Keywords: Language Modeling, Generation, Dialogue and Interactive Systems

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English, Chinese

Submission Number: 8694