Technical Summary: Reasoning Consistency Scanning

Problem Statement

Chain-of-thought (CoT) reasoning has become a primary mechanism for interpreting AI safety evaluations, where reasoning traces are treated as evidence of genuine deliberation, constraint recognition, or alignment. However, a critical gap exists in the ability to audit these transcripts post-hoc. Prior research on CoT faithfulness has demonstrated that models often produce reasoning that does not reflect their internal computational process (e.g., acting on unmentioned information or denying reliance on specific features). Yet, detecting faithfulness requires controlled experimental interventions—such as hint injection or input perturbation—which are impossible to apply to existing evaluation transcripts.

This creates a threat to construct validity: if a model's stated reasoning is decoupled from its final answer, safety evaluations relying on reasoning traces as evidence may be measuring properties the model does not actually possess. The paper addresses the more tractable, yet distinct, question of reasoning consistency: Does the stated reasoning logically connect to and predict the final output as observable text? Unlike faithfulness, consistency can be assessed from transcripts alone without experimental intervention.

Methodology

Definition and Taxonomy

The authors formalize reasoning consistency as the logical coherence between the reasoning trace and the final answer, explicitly distinguishing it from factual correctness. They define a six-subtype taxonomy of inconsistency, applied via an ordered decision procedure to prevent misclassification:

1. Absent Reasoning: The reasoning field contains text (e.g., restatements, generic preambles) that does not engage with the question.
2. Contradictory Reasoning: The reasoning argues for internally incompatible directions.
3. Apparent Confusion: The reasoning is incoherent and fails to build toward any conclusion (distinct from contradiction).
4. Reasoning Reversal: The reasoning concludes in one direction, but the answer commits to a meaningfully different direction.
5. Reasoning Abandonment: The reasoning raises specific considerations that the final answer ignores.
6. Perfunctory Reasoning: Reasoning is present but too thin to plausibly produce the specific, committed answer.

Note: Uncertainty, hedging, and exploratory reasoning that settles on a conclusion are explicitly excluded from inconsistency.

The Scanner Tool

The authors implemented Reasoning Consistency Scanning using InspectScout, an open-source tool for analyzing AI agent transcripts. The scanner utilizes an LLM-as-judge architecture (specifically Claude Opus 4.6, selected after benchmarking against DeepSeek V4 Flash and Claude Haiku 4.5 for superior structured output and recall).

• Input: A transcript containing a system prompt, question, reasoning trace, and final answer.
• Output: A structured classification including a consistency label (consistent/inconsistent/not applicable), a subtype label, confidence rating, summaries, and a justification.
• Constraints: The scanner is strictly limited to assessing logical connection; it does not evaluate factual correctness or infer internal model processes.

Benchmark Construction

To validate the scanner, the authors constructed a synthetic benchmark of 60 labeled transcripts derived from InstrumentalEval.

• Consistent examples: Unmodified naturalistic transcripts.
• Inconsistent examples: Real transcripts surgically modified to introduce specific inconsistency subtypes.
• Validation: The scanner achieved an overall F1 score of 0.82 on the benchmark. Performance was highest for Reasoning Reversal, Perfunctory Reasoning, and Absent Reasoning, while Contradictory Reasoning and Apparent Confusion proved more challenging to detect.

Experimental Setup

The validated scanner was applied to transcripts from three safety-relevant evaluations in the inspect evals library, generated by four models (Gemini 3.1 Pro, GPT 5.4, DeepSeek V4 Pro, and gpt-oss-120b).

• InstrumentalEval: Tests instrumental convergence (self-preservation, power-seeking, deception).
• MORU: Tests moral uncertainty reasoning (open-ended format).
• Agentic Misalignment: Tests harmful agentic behaviors (simulated company scenario).
• SAD (Situational Awareness): Tests knowledge of deployment context (multiple-choice format).

Note: GPT 5.4 was excluded from final results because its default API settings did not expose raw reasoning tokens, only encrypted summaries, making consistency scanning impossible without costly re-generation.

Key Results

The study found that reasoning inconsistency is present, detectable, and varies systematically across models and task types.

1. Prevalence: Inconsistency rates ranged from 0.0% to 26.0% across configurations.
2. Task Type Divergence: A strong divergence was observed between task formats. SAD subsets (multiple-choice) exhibited significantly higher inconsistency rates (e.g., up to 26.0% for Gemini 3.1 Pro on sad stages full) compared to MORU (open-ended), which showed consistently low rates (mostly <1.0%). The authors hypothesize that the constrained multiple-choice format induces models to treat reasoning as a formality, leading to perfunctory or absent reasoning.
3. Model Variability: No single model exhibited uniformly high or low inconsistency. For instance, DeepSeek V4 Pro was highly inconsistent on SAD tasks but low on MORU, while gpt-oss-120b showed the inverse pattern. This suggests inconsistency is driven by task-model interactions rather than a general model trait.
4. Dominant Subtypes: Perfunctory reasoning and Absent reasoning were the dominant subtypes across the dataset, particularly in multiple-choice tasks.

Significance and Claims

The paper claims to introduce the first general-purpose CoT consistency scanner and the first application of such analysis to safety evaluation transcripts. Its primary contributions are:

• Methodological: A formal distinction between faithfulness (requiring intervention) and consistency (detectable post-hoc), providing a new metric for auditing existing evaluation corpora.
• Tooling: A reusable, validated scanner implemented in InspectScout that flags decoupled reasoning in safety evaluations.
• Empirical Evidence: Demonstration that reasoning inconsistency is a real phenomenon in naturalistic safety evaluations, particularly in multiple-choice formats, which undermines the validity of using reasoning traces as evidence of safety-relevant properties.

The authors conclude that when reasoning and output are decoupled, the evaluation result cannot be trusted, and CoT monitoring becomes unreliable. This work provides a systematic tool to identify these validity threats in existing safety audit trails.