Evaluating Stochastic Collapse and Implicit Bias in Multimodal Large Language Models

A new benchmark reveals that leading multimodal AI models consistently fail to produce random outputs when instructed, instead collapsing toward predictable favorites even in scenarios where all choices are equally valid, according to research posted on arXiv [1]. The study, submitted June 4, introduces RandomBench, a tool to evaluate whether Multimodal Large Language Models (MLLMs) can maintain distributionally neutral behavior when selecting among equivalent options [1]. Current evaluations for these models overwhelmingly focus on utility-driven objectives, leaving behavior in logic-neutral scenarios largely underexplored [2]. Stochasticity is essential in settings where multiple actions are equally valid, such as recommending travel itineraries or daily schedules where several options carry similar utility [2]. Deterministic policies in such contexts may lead to repetitive behaviors and reduced coverage of valid alternatives [2]. The researchers documented a phenomenon they call Stochastic Collapse, where MLLMs fail to maintain uniform randomness despite explicit random instructions [1]. In tests, top-1 probabilities reached 97% from the ideal one-quarter baseline, and the Randomness Index dropped to 0.068 in Claude Sonnet 4.6 [2]. Claude is a series of large language models developed by Anthropic, an American AI company founded in 2021 with a focus on AI safety [6]. Large language models are machine learning models with many parameters, trained with self-supervised learning on vast amounts of text [7]. The benchmark introduces three metrics—RI, BCI, and BII—to quantify entropy and distributional bias [2]. Ablation studies demonstrated that these deviations persist across languages and representation formats, underscoring the robustness of distributional collapse in logic-neutral decision settings [2]. The findings arrive as the AI field continues to grapple with evaluation standards. Popular benchmarks such as MMLU have long been used to assess language model capabilities, inspiring spin-offs like MMLU-Pro and MMLU-Redux [8]. Machine learning, the broader discipline underpinning these models, relies on statistical algorithms that learn from pre-trained data and generalize to unseen tasks without explicit programming [4]. Deep learning, a subset of machine learning, uses multilayered neural networks to perform classification, regression, and representation learning [5]. The RandomBench results suggest that even advanced architectures exhibit systematic biases when asked to behave randomly, a finding with implications for applications ranging from content recommendation to scheduling systems where variety and fairness depend on genuine stochasticity.