Emergent Bayesian Behaviour and Optimal Cue Combination in LLMs

ICLR 2026 Conference Submission25197 Authors

Keywords: Large Language Models (LLMs), Psychophysics, Bayesian Inference, Cue Combination, Emergent Abilities, LLM Evaluation, Uncertainty Quantification

Abstract:

Large language models (LLMs) excel at explicit reasoning, but their implicit computational strategies remain underexplored.
Decades of psychophysics research show that humans intuitively process and integrate noisy signals using near-optimal Bayesian strategies in perceptual tasks. We ask whether LLMs exhibit similar behaviour and perform optimal multimodal integration without explicit training or instruction. Adopting the psychophysics paradigm, we infer computational principles of LLMs from systematic behavioural studies. We introduce a behavioural benchmark - BayesBench: four magnitude-estimation tasks (length, location, distance, and duration) over text and image, inspired by classic psychophysics, and evaluate a diverse set of nine LLMs alongside human judgements for calibration. Through controlled ablations of noise, context, and instruction prompts, we measure performance, behaviour and efficiency in multimodal cue-combination. Beyond accuracy and efficiency metrics, we introduce a Bayesian Consistency Score that detects Bayes-consistent behavioural shifts even when accuracy saturates. Our results show that high task accuracy - notably for GPT-5 Mini - does not always imply efficient cue combination; yet accurate models, including GPT-5 Mini, Llama 4 Maverick, and Claude 3.7 Sonnet, often adapt in Bayes-consistent ways. These findings reveal emergent principled handling of uncertainty and highlight the correlation between accuracy and Bayesian tendencies. We release our psychophysics benchmark and consistency metric as evaluation tools and to inform future multimodal architecture designs.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 25197