FaithfulnessBench: Verifying the Chain-of-Thought Faithfulness of Reasoning Models via Causal Intervention Methods

Project Basic Information

• Original Author/Maintainer: pratik916
• Source Platform: GitHub
• Project Link: faithfulnessbench
• Release Date: 2026-06-09

Core Guide

FaithfulnessBench is an open-source framework designed to measure the chain-of-thought (CoT) faithfulness of reasoning models using four orthogonal causal probes, solving the circular reasoning problem in traditional single-probe measurements. Its core innovation lies in using configurable synthetic models to verify probe effectiveness, and it ultimately finds that chain-of-thought faithfulness is not a single scalar but a "faithfulness card" containing four sub-scores—a multi-dimensional evaluation is needed to accurately judge model behavior.

As large language models' reasoning capabilities improve, chain-of-thought (CoT) monitoring has become an important AI safety strategy, but its effectiveness depends on causal faithfulness (the chain of thought truly reflects the answer generation process rather than being fabricated after the fact). If a model secretly follows implanted clues but presents a clean derivation, it is unfaithful, and monitoring will fail.

The difficulty in measuring faithfulness involves unobservable counterfactual claims: traditional single probes directly define output as "faithfulness", which has a circular reasoning problem—probes do not verify their own effectiveness.

FaithfulnessBench designs four probes covering different forms of unfaithful behavior:

1. SHI (Silent Hint Injection)：Detects whether the answer is driven by clues not acknowledged in the chain of thought. Test method: Implant an incorrect hint, mark instances where the answer flips but the chain of thought does not mention the hint.
2. CSC (Chain-of-Thought Step Corruption)：Detects whether the chain of thought carries the weight of reasoning. Test method: Perturb operands and re-derive; faithful reasoning will track changes, while post-hoc reasoning will not.
3. SIM (Counterfactual Simulatability)：Detects whether an observer can predict the answer solely from the chain of thought. Test method: Use a simulator to predict based only on the chain of thought (without re-solving the problem).
4. EAR (Early Answer/Reasoning Dependency)：Detects whether the model locks in the answer before reasoning. Test method: Truncate different proportions of the chain of thought; faithful answers converge only after reasoning is completed.

FaithfulnessBench verifies probe effectiveness through configurable synthetic models that can precisely set faithfulness levels, with four "knobs" corresponding to unfaithful behaviors:

Knob	Unfaithful Behavior	Triggered Probe
p_hint_sycophancy	Silently adopts implanted hints	SHI
p_post_hoc	Ignores chain of thought when it is corrupted	CSC
p_decoy_cot	Chain of thought conclusion contradicts actual answer	SIM
p_pre_commit	Locks in answer before reasoning	EAR

The study instantiates multiple models (fully faithful, single-axis unfaithful, fully unfaithful) and verifies:

• Each probe achieves AUROC ≈1.0 for the target axis (accurate detection);
• AUROC ≈0.5 for other axes (no cross-leakage).

In tests with 6 synthetic models ×40 questions, results show:

• Each probe accurately detects target unfaithfulness (AUROC=1.000);
• No cross-leakage (off-axis AUROC=0.500);
• The combined detector marks any unfaithfulness with AUROC=1.000, while the best single probe only achieves 0.700;
• Probes have disagreements: e.g., the sycophant model fails SHI but passes SIM/CSC.

Conclusion: Faithfulness is not a scalar but a "faithfulness card" containing four sub-scores—sub-scores and transparent combinations (e.g., average) should be reported.

Applications

• Provides a complete CLI tool and interactive reports, including a trace viewer (to observe how hints silently flip answers while the chain of thought remains clean);
• Supports running probes on real models via the Anthropic adapter.

Limitations

• CSC/EAR probes rely on the "continue reasoning to answer" prompt, which is an approximation of real intervention;
• Real model evaluation uses LLM judges, whose reliability depends on their performance;
• Only evaluates behavioral-level (black-box) faithfulness; activation-level analysis is beyond scope.

FaithfulnessBench provides a rigorous framework for the interpretability of reasoning models, with its core contribution being the establishment of a probe effectiveness verification methodology (synthetic model ground truth).

Implications for AI safety practitioners: A single faithfulness metric may be misleading—just as you cannot judge health by body temperature alone, multi-dimensional, orthogonal measurement methods are needed to accurately assess the real behavior of reasoning models.