Superficial Beliefs in LLM Decision-Making

Superficial Beliefs in LLM Decision-Making, from Imperial College London, asks whether models such as OpenAI’s GPT-5-mini and GPT-5-nano, Qwen3-14B, and Ministral-3-14B merely rehearse explanations or instead exhibit a recoverable local decision structure. The authors build a synthetic binary-choice benchmark with four graded attributes across drugs, policy, and software themes, plus two control themes and a structurally different six-attribute hospital cyber-response setting. They fit a binomial logistic behavioural model to 400 training families and evaluate on 100 held-out families, comparing its recovered driver with two explicit elicitation methods: direct response and a score-based judge. Across the main benchmark, the behavioural model matches held-out choices 80.4% of the time, while the explicit direct report names the recovered driver only 61.0% of the time and the score-based judge 61.3% of the time. The same pattern survives perturbation and intervention: score-based judge outputs are more reproducible than direct responses, yet they are not more faithful to the recovered decision driver, and in the control themes irrelevant attributes are selected only 0.2%–1.6% of the time. A targeted occlusion study on GPT-5-mini non-thinking drugs shows that higher-ranked attributes induce larger choice and rationale flips, supporting the claim that these models behave as if guided by probabilistic, attribute-local priorities rather than fully articulated beliefs.