Representational Control over Self-Report & Behavior Coherence in LLM Risk-Taking
Rafal Kocielnik, Peiyang Song, Pengrui Han, Myrl G. Marmarelis, Ramit Debnath, Dean Mobbs, R. Michael Alvarez · arXiv (2026) · doi:10.48550/arxiv.2610.04125 · gemini-3.1-flash-lite
TL;DR
Researchers tested whether an AI's claimed personality about taking risks matches its actual choices in gambling games. They found that these two behaviors are controlled by different parts of the AI's internal logic, meaning the AI can be made to say it is cautious while acting recklessly. These results suggest that asking an AI about its intentions is not a reliable way to predict what it will actually do.
Method and sample: The study tested four open-weight LLMs (Qwen2.5-7B, Llama3.1-8B, OLMo-3-7B, and Ministral-3-8B). Risk-taking was measured via the Columbia Card Task (CCT) and Balloon Analogue Risk Task (BART), while self-report was assessed via BFI-44 and DOSPERT. The study analyzed nine steering-vector extraction methods spanning task directives, model choices, and personality descriptions.
#mechanistic-interpretability#activation-steering#assessment-validity#risk-taking#representation-engineering