C2-Faith: Benchmarking LLM Judges for Causal and Coverage Faithfulness in Chain-of-Thought Reasoning

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

C2-Faith, a new benchmark built from PRM800K, evaluates Large Language Model (LLM) judges' ability to assess chain-of-thought (CoT) reasoning faithfulness. It explicitly decomposes faithfulness into causality, ensuring logical step progression, and coverage, verifying essential intermediate inferences. The benchmark uses controlled perturbations, replacing steps for known causal errors and deleting portions for coverage errors, allowing direct measurement against reference labels. Evaluating three frontier LLM judges across binary causal detection, causal step localization, and coverage scoring tasks, results show judge reliability is highly task-dependent. Models detect errors but struggle with accurate localization. Furthermore, judges systematically overestimate reasoning completeness, assigning high scores even with significant missing intermediate steps. These findings expose fundamental limitations in process-level LLM evaluation.

Key takeaway

For NLP Engineers evaluating LLM-generated chain-of-thought reasoning, recognize that current LLM judges often misattribute errors and inflate completeness scores. You should implement additional verification steps beyond simple binary detection, focusing on methods that can precisely localize reasoning flaws. Consider human-in-the-loop review for critical applications to ensure accurate process-level evaluation and avoid over-reliance on automated LLM assessments.

Key insights

LLM judges struggle with precise error localization and systematically overestimate reasoning completeness in CoT evaluations.

Principles

• Faithfulness decomposes into causality and coverage dimensions.
• LLM judge reliability is highly task-dependent across evaluation settings.
• Error detection differs significantly from accurate error attribution.

Method

C2-Faith uses controlled perturbations on PRM800K, replacing steps for causal errors and deleting portions for coverage errors, enabling direct measurement against reference labels.

In practice

• LLMs over-assign high coverage scores to incomplete reasoning.
• LLMs detect CoT errors but struggle to localize them accurately.

Topics

• LLM Benchmarking
• Chain-of-Thought Reasoning
• Causal Faithfulness
• Coverage Faithfulness
• LLM Judges
• Reasoning Evaluation

Best for: Research Scientist, AI Engineer, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Mechanistic Evidence for Faithfulness Decay in Chain-of-Thought Reasoning — NLDD quantifies CoT faithfulness, revealing models can achieve high accuracy without causally relying on their generated reasoning.
• From Benchmarking to Reasoning: A Dual-Aspect, Large-Scale Evaluation of LLMs on Vietnamese Legal Text — LLMs struggle with legal reasoning and application, often masking critical errors despite high surface-level accuracy.
• Do LLMs Game Formalization? Evaluating Faithfulness in Logical Reasoning — LLMs can produce unfaithful logical proofs despite high compilation rates, highlighting a gap between validity and faithfulness.
• When the Judge Is Wrong — LLM-as-judge is unreliable for structured tasks, even with ground truth or source documents.
• Know Your Limits : On the Faithfulness of LLMs as Solvers and Autoformalizers in Legal Reasoning — LLMs achieve high accuracy in legal reasoning benchmarks but often lack true logical faithfulness, exhibiting critical failure modes.
• Breaking the Chain: A Causal Analysis of LLM Faithfulness to Intermediate Structures — LLMs' apparent faithfulness to intermediate reasoning structures is fragile; they often fail to update predictions after causal intervention.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.