Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note2 min read

Reasoning Consistency Scanning: A Framework for Auditing Chain-of-Thought Validity in AI Safety Evaluations

Source

arXiv

arxiv.org/abs/2607.07229v1

Read the source

Source

Plain-Language Brief

The paper introduces a method called Reasoning Consistency Scanning to assess the logical consistency between stated reasoning and answers in AI safety evaluation transcripts. The method can be applied to various AI generator models and evaluation tasks, and the authors provide a benchmark of 60 transcripts for validation.

What the Document Says

  • The paper presents a framework for evaluating the logical consistency of stated reasoning in AI safety evaluation transcripts.
  • A six-subtype taxonomy of inconsistency is defined.
  • A benchmark of 60 transcripts is built for validation.
  • An implementation of the scanner for InspectScout is presented.
  • The authors report results across four generator models and three evaluations, showing that reasoning inconsistency exists, can be detected, and varies systematically.

Key Concepts Mentioned

  • AI safety evaluation
  • chain-of-thought (CoT) reasoning
  • logical consistency
  • reasoning consistency scanning
  • InspectScout
  • benchmark

Claims Made by the Document

  • Prior work shows that chain-of-thought reasoning is often unfaithful.
  • Detecting unfaithfulness requires controlled experimental interventions, which cannot be applied to evaluation transcripts after the fact.
  • Consistency can be assessed from a transcript alone without intervention.
  • The introduced method can detect logical consistency in AI safety evaluation transcripts.
  • The benchmark provides 60 manually adapted transcripts from InstrumentalEval outputs.

Evidence Used

  • The paper cites prior work on chain-of-thought reasoning.
  • The authors explain that detecting unfaithfulness requires controlled experimental interventions.
  • The authors state that consistency can be assessed from a transcript alone without intervention.
  • The authors present the method and its implementation.
  • The authors mention the benchmark of 60 manually adapted transcripts.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Prior work has shown that chain-of-thought (CoT) reasoning is often unfaithful: a model's stated reasoning does not reliably reflect the process that produced its output. Detecting unfaithfulness, though, requires controlled experimental interventions, which cannot be applied to evaluation transcripts after the fact. We turn instead to a more tractable question that has received less attention: whether the stated reasoning is logically consistent with the answer it accompanies. Unlike faithfulness, consistency can be assessed from a transcript alone, with no intervention. We introduce reasoning consistency scanning, a reusable method for detecting this property in AI safety evaluation transcripts. Our contributions are fourfold. First, we formalize reasoning consistency as distinct from faithfulness and define a six-subtype taxonomy of inconsistency. Second, we build a validated benchmark of 60 transcripts, manually adapted from InstrumentalEval outputs. Third, we implement a working scanner for InspectScout, the first to target this property in safety evaluation transcripts. Fourth, we report results across four generator models and three evaluations from inspect_evals, showing that reasoning inconsistency is present, detectable, and varies systematically across both models and task types.