Research map / Large language models / Reasoning and chain of thought

FaithCoT-Bench: Benchmarking Instance-Level Faithfulness of Chain-of-Thought Reasoning

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces FaithCoT-Bench, a benchmark with an expert-annotated dataset (FaithCoT-Bench dataset, over 1,000 CoT traces from four LLMs across four domains) for instance-level CoT unfaithfulness detection, and benchmarks 11 detection methods.

Problem

Existing CoT faithfulness studies are mechanism-level and aggregate; there is no expert-verified instance-level ground truth or systematic evaluation for deciding whether a specific CoT trace is unfaithful.

Approach

Formulates instance-level CoT unfaithfulness detection as binary classification, builds an expert-annotated dataset with unfaithfulness labels, reasons, and step-level evidence under a two-category taxonomy (post-hoc reasoning, spurious reasoning chains), and benchmarks counterfactual, logit-based, and LLM-as-judge detectors.

Key results

• LLM-as-judge methods (Faithful-Judge) achieve the best F1 in most settings (e.g., 77.7 on LogicQA with LLaMA3.1), while logit-based methods often score below 50.
• Detection is harder in knowledge-intensive domains: on HLE-Bio, Information Gain drops to 9.5 for LLaMA3.1 and Faithful-Judge scores fall to ~50 for GPT-4o-mini and Gemini on TruthfulQA.
• Over 1,000 trajectories collected, with more than 300 annotated unfaithful; inter-annotator Cohen's kappa 81.0–97.2.

Surprising findings

• Stronger models (GPT-4o-mini, Gemini 2.5 Flash) do not yield easier faithfulness detection; detector F1 is often lower on their outputs.
• Higher task accuracy does not imply higher faithfulness: Qwen2.5-7B has higher AQuA accuracy but more unfaithful CoTs.

Limitations

• Labels rely on observable signals and human judgment because the true internal reasoning path is unobservable.
• Only four generator models and 378 queries are used.
• Several result descriptions and examples are internally inconsistent, reducing clarity.

Experimental setup

Datasets: LogicQA TruthfulQA AQuA HLE-Bio

Benchmarks: LogicQA TruthfulQA AQuA HLE-Bio

Metrics: F1 accuracy Cohen's kappa

Baselines: random classifier perplexity adding mistakes option shuffling removing steps early answering paraphrasing answer tracing information gain Step-Judge Faithful-Judge

Builds on: chain-of-thought prompting

Closest papers by meaning

• Deep Hidden Cognition Facilitates Reliable Chain-of-Thought Reasoning (AAAI 2026 · similarity 0.92)
• Unlocking the Capabilities of Thought: A Reasoning Boundary Framework to Quantify and Optimize Chain-of-Thought (NeurIPS 2024 · similarity 0.91)
• The CoT Encyclopedia: Analyzing, Predicting, and Controlling how a Reasoning Model will Think (ICLR 2026 · similarity 0.90)
• How Does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding (AAAI 2026 · similarity 0.90)
• Long Chain-of-Thought Reasoning Across Languages (ICLR 2026 · similarity 0.90)
• Causal Sufficiency and Necessity Improves Chain-of-Thought Reasoning (NeurIPS 2025 · similarity 0.90)