Fudan Finds RandomBench Exposes How GPT and Claude Almost Never Choose Randomly
A new benchmark exposes that frontier multimodal AI models cannot make truly random choices, collapsing to a single option up to 97% of the time even when explicitly told to pick randomly.
- Stochastic Collapse discovered: Frontier multimodal LLMs fail to pick randomly among equivalent options, with top-1 choice probability reaching 97% (ideal: 25%).
- RandomBench introduced: A 200-instance benchmark across text and vision modalities, using 50 repeated samples per instance to expose distributional bias.
- Three new metrics: RI (Randomness Index), BCI (Bias Consistency Index), and BII (Bias Intensity Index) quantify how far models deviate from uniform random behavior.
- All 7 frontier models fail: GPT-5.1, Claude Sonnet 4.6, Gemini 3.1, Grok 4, Qwen 3.6, Kimi K2.5, and Doubao Seed 1.6 all exhibit collapse; Claude Sonnet 4.6 scores RI of 0.068 out of 1.0.
- Visual hijacking identified: In vision tasks, perceptual cues like brightness, spatial position, and geometric complexity override random-choice instructions.
- Bias persists across languages and formats: Ablations show collapse survives prompt language changes and label format changes, indicating it is deeply embedded in model training.
Ask a frontier AI model to flip a coin and pick randomly between four equally valid options. You might expect it to spread its choices roughly evenly. What researchers from Fudan University and Beihang University found instead is alarming: the models almost always pick the same answer. They named this failure mode Stochastic Collapse, and it turns out it affects every major multimodal model tested, from GPT to Claude to Gemini.
Why randomness actually matters
Stochasticity is essential in scenarios where multiple actions are equally valid, such as recommending travel itineraries or daily schedules. In such settings, deterministic policies may lead to repetitive behaviors and reduced coverage of valid alternatives. This is not a niche edge case. Any agent that needs to explore options, diversify recommendations, or simulate fair sampling runs into this problem. Large language models can often accurately describe probability distributions using natural language, yet they still struggle to generate faithful samples from them -- a mismatch that limits their use in tasks requiring reliable stochasticity, such as Monte Carlo methods, agent-based simulations, and randomized decision-making.
The broader bias landscape for LLMs is well-documented. Prior work has shown that LLMs exhibit selection bias when answering multiple-choice questions, often preferring specific option identifiers such as "Option A" due to token-level probability biases. This bias frequently manifests as positional preferences, where models disproportionately select answers appearing in certain positions. What this new paper adds is a systematic, multimodal treatment of the problem -- and the discovery that the collapse is far more severe than previously understood.
The benchmark: 200 logic-neutral problems
To bridge this critical gap, the researchers introduce RandomBench, designed to quantify model behavior under absolute semantic equivalence. RandomBench features 200 instances across two modalities: RB-Text and RB-Vision. The key design principle is logic-neutrality -- every question in the dataset has options that are genuinely equivalent, so there is no correct answer. There is no utility difference between the choices. The only right behavior is a uniform distribution.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.