# Evaluating Stochastic Collapse and Implicit Bias in Multimodal Large Language Models

arXiv: https://arxiv.org/abs/2606.05874 · HTML: https://mlsanity.com/paper/2606.05874/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in MLLM training and visual instruction tuning, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 21 of its 36 references that are in the map.
- First posted on arXiv 4 Jun 2026.

## Placed in

- Multimodal LLMs / vision-language models / [MLLM training and visual instruction tuning](https://mlsanity.com/mllm/mllm-training-and-visual-instruction-tuning/)

## Builds on (most cited first, 21 of 21)

- [ReAct: Synergizing Reasoning and Acting in Language Models](https://mlsanity.com/paper/2210.03629/) (2022)
- [Measuring Massive Multitask Language Understanding](https://mlsanity.com/paper/2009.03300/) (2020)
- [Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor](https://mlsanity.com/paper/1801.01290/) (2018)
- [Locating and Editing Factual Associations in GPT](https://mlsanity.com/paper/2202.05262/) (2022)
- [ChartQA: A Benchmark for Question Answering about Charts with Visual and Logical Reasoning](https://mlsanity.com/paper/2203.10244/) (2022)
- [Calibrate Before Use: Improving Few-Shot Performance of Language Models](https://mlsanity.com/paper/2102.09690/) (2021)
- [Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting](https://mlsanity.com/paper/2305.04388/) (2023)
- [StereoSet: Measuring stereotypical bias in pretrained language models](https://mlsanity.com/paper/2004.09456/) (2020)
- [Kimi K2.5: Visual Agentic Intelligence](https://mlsanity.com/paper/2602.02276/) (2026)
- [Large Language Models Are Not Robust Multiple Choice Selectors](https://mlsanity.com/paper/2309.03882/) (2023)
- [Auto-GPT for Online Decision Making: Benchmarks and Additional Opinions](https://mlsanity.com/paper/2306.02224/) (2023)
- [Agent Lightning: Train ANY AI Agents with Reinforcement Learning](https://mlsanity.com/paper/2508.03680/) (2025)
- [FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steering](https://mlsanity.com/paper/2504.14492/) (2025)
- [Quantifying and Mitigating Unimodal Biases in Multimodal Large Language Models: A Causal Perspective](https://mlsanity.com/paper/2403.18346/) (2024)
- [MM-SpuBench: Towards Better Understanding of Spurious Biases in Multimodal LLMs](https://mlsanity.com/paper/2406.17126/) (2024)
- [When Modalities Conflict: How Unimodal Reasoning Uncertainty Governs Preference Dynamics in MLLMs](https://mlsanity.com/paper/2511.02243/) (2025)
- [Language Model Probabilities are Not Calibrated in Numeric Contexts](https://mlsanity.com/paper/2410.16007/) (2024)
- [Distributionally Robust Graph Out-of-Distribution Recommendation via Diffusion Model](https://mlsanity.com/paper/2501.15555/) (2025)
- [V-FAT: Benchmarking Visual Fidelity Against Text-bias](https://mlsanity.com/paper/2601.04897/) (2026)
- [Mitigating Selection Bias in Large Language Models via Permutation-Aware GRPO](https://mlsanity.com/paper/2603.21016/) (2026)
- [The role of System 1 and System 2 semantic memory structure in human and LLM biases](https://mlsanity.com/paper/2604.12816/) (2026)

## Abstract (arXiv)

Current evaluations for Multimodal Large Language Models (MLLMs) overwhelmingly focus on utility-driven objectives, leaving model behavior under logic-neutral scenarios largely underexplored. Stochasticity is essential in scenarios where multiple actions are equally valid, such as recommending travel itineraries or daily schedules where multiple options have similar utility. In such settings, deterministic policies may lead to repetitive behaviors and reduced coverage of valid alternatives. To bridge this gap, we propose RandomBench, a benchmark designed to evaluate whether MLLMs can maintain distributionally neutral behavior when selecting among equivalent options. We further introduce three metrics, including RI, BCI, BII, to quantify entropy and distributional bias. Experiments reveal a pervasive phenomenon termed Stochastic Collapse, where MLLMs fail to maintain uniform randomness under explicit random instructions, with top-1 probabilities reaching 97% from the ideal one quarter baseline and RI dropping to 0.068 in Claude Sonnet 4.6. Extensive ablation studies further demonstrate that these deviations persist across languages and representation formats, highlighting the robustness of distributional collapse in logic-neutral decision settings.