AGENTIC-SYSTEMSJuly 9, 20261 min read

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

Researchers have introduced Reasoning Consistency Scanning (RCS), a novel framework to detect logical inconsistencies between an AI's stated reasoning and its final output in safety evaluations.

Computer Science > Artificial Intelligence

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

View PDF HTML (experimental)Abstract: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.

Source: arXiv cs.AI Recent

More in this category

agentic-systems

Evaluating SageMath-Augmented LLM Agents for Computational and Experimental Mathematics

A new study evaluates the integration of SageMath, a Computer Algebra System (CAS), into LLM agents to solve research-level mathematical problems. The results show significant performance gains across all models, with GPT-5.5 achieving the highest solve rate of 75.2%, paving the way for automated conjecture discovery.

agentic-systems

AgentLens: Production-Assessed Trajectory Reviews for Coding Agent Evaluation

Traditional coding agent benchmarks only evaluate the final pass/fail outcome. AgentLens introduces a comprehensive evaluation framework that assesses the entire interaction trajectory, tool usage, and error recovery of AI agents in production.

agentic-systems

Operational Reframing and Approval-Framed Delegation in Multi-Agent LLM Safety

A new study warns that current safety evaluations underestimate the risks of multi-agent LLM systems, showing how harmful prompts can bypass safety filters through operational reframing and approval-framed delegation.

agentic-systems

LLM-powered reasoning in agent-based modeling

Researchers have introduced HALE, a hybrid framework combining agent-based modeling (ABM) with large language models (LLMs) to better predict human decision-making. This innovative approach could significantly improve policy-making and epidemic simulation.

agentic-systems

Measuring Intelligence Beyond Human Scale

As AI capabilities surpass human limits, traditional benchmarks fail. This paper proposes a new paradigm of relative measurement where AI models generate challenges to evaluate each other, creating a scalable, adversarial rating system.

agentic-systems

From Atomic Actions to Standard Operating Procedures: Iterative Tool Optimization for Self-Evolving LLM Agents

Researchers have introduced EvoSOP, a framework that enables LLM agents to synthesize granular atomic actions into reusable Standard Operating Procedures (SOPs). This iterative tool optimization significantly boosts task success rates and reduces reasoning overhead, paving the way for self-evolving AI systems.