New framework audits AI reasoning consistency in safety evaluations

Researchers have introduced a new framework called Reasoning Consistency Scanning to audit the validity of Chain-of-Thought (CoT) reasoning in AI safety evaluations. This method focuses on detecting logical inconsistencies within a model's stated reasoning and its accompanying answer, a property that can be assessed from transcripts alone. The framework includes a formalized definition of reasoning consistency, a benchmark of 60 transcripts adapted from InstrumentalEval, and a scanner implemented for InspectScout. Initial results across four generator models and three evaluations indicate that reasoning inconsistency is detectable and varies systematically. AI

IMPACT This framework could improve the reliability of AI safety evaluations by ensuring that the reasoning processes of models are logically sound and consistent.

RANK_REASON The item is an academic paper detailing a new framework and benchmark for auditing AI reasoning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• inspect_evals
• InspectScout
• InstrumentalEval
• Reasoning Consistency Scanning
• ScienceCast

• paper
• safety

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →