💬 NLP

Evaluating Stochastic Collapse and Implicit Bias in Multimodal Large Language Models

This paper introduces RandomBench, a benchmark and a set of metrics to evaluate Multimodal Large Language Models' ability to maintain distributionally neutral behavior in logic-neutral scenarios, revealing a pervasive phenomenon called "Stochastic Collapse" where models exhibit strong implicit bias and fail to generate uniform randomness even when explicitly instructed to do so.

Original authors: Huiyuan Zheng, Houtao Zhang, Boyang Wang, Qingyi Si, Hongcheng Guo

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Huiyuan Zheng, Houtao Zhang, Boyang Wang, Qingyi Si, Hongcheng Guo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The "Unfair Coin" Problem

Imagine you ask a friend to flip a coin and tell you if it landed on Heads or Tails. If they are truly random, they should pick Heads 50% of the time and Tails 50% of the time.

Now, imagine you ask a super-smart AI (a Multimodal Large Language Model or MLLM) to do the exact same thing: "Flip a coin. Tell me Heads or Tails." You expect the AI to be perfectly fair.

The paper's shocking discovery: The AI's "coin" is not fair. It is heavily weighted. Even when the AI is told to be random, it secretly has a favorite side. It might pick "Heads" 97% of the time and "Tails" only 3% of the time, even though it knows it's supposed to be random.

The authors call this "Stochastic Collapse." It's like the AI's brain gets stuck in a rut, unable to truly let go and choose randomly.

The Tool: "RandomBench" (The Test Kitchen)

To prove this, the researchers built a new test called RandomBench. Think of this as a "logic-free zone" or a "neutral playground."

The Goal: Create situations where there is no right answer.
The Setup: They created 200 different scenarios. Some were just text (e.g., "Pick a letter: A, B, C, or D"), and some involved pictures (e.g., "Pick one of these four identical-looking shapes").
The Rule: Every option was exactly equal. There was no reason to pick A over B. If the AI was truly random, it would pick each option 25% of the time.

They asked 7 different top-tier AI models to play this game 50 times for every single scenario. That's 10,000 total "coin flips."

What They Found: The AI's "Secret Habits"

The results showed that the AI models failed the test spectacularly. Instead of being a fair coin, they acted like a broken slot machine that always lands on the same symbol.

Here are the main "bad habits" they found, explained with metaphors:

1. The "Visual Hijack" (The Loud Sign)

When the AI was shown pictures, its "randomness" broke even faster.

The Metaphor: Imagine you are asked to pick a door at random. But one door has a tiny, almost invisible smudge on it, while the others are clean. Even though you are told to ignore the smudge, your brain is drawn to it.
The Finding: The AI couldn't ignore visual details. If one option was slightly brighter, in a specific corner, or had a tiny blur, the AI would pick that one almost every time. The visual cue "hijacked" the instruction to be random.

2. The "Positional Bias" (The Middle Seat)

The Metaphor: Imagine a row of four empty chairs. You are told to sit in a random one. Humans might pick the middle one because it feels "safe" or "central."
The Finding: The AI had a strong preference for specific spots. It loved the center, the top, or the right side. Even if the options were just four blank squares, the AI would pick the top-right one 90% of the time.

3. The "Language Switch" (The Accent Effect)

The Metaphor: Imagine you ask a person to pick a random number in English, and they pick "7." Then you ask the same person to pick a random number in French, and they suddenly switch to picking "3."
The Finding: The AI's "favorite" choice changed depending on the language used. The same visual image would lead to a totally different "random" choice if the instructions were in Chinese vs. English. This proves the bias is deeply tied to the language the AI was trained on.

4. The "Stronger is Worse" Paradox

The Metaphor: You might think a smarter, more obedient robot would be better at following the "be random" rule.
The Finding: Surprisingly, the smarter and more "aligned" models (those that follow instructions very well) were actually worse at being random. They were so good at finding a pattern to justify their choice that they locked onto a single option and stuck with it, often making up a logical-sounding excuse for why they picked it, even though there was no logic to begin with.

Why Does This Happen? (The "System 1" Glitch)

The paper uses a concept from psychology called Dual Process Theory:

System 2: Slow, logical thinking (like solving a math problem).
System 1: Fast, intuitive, gut-feeling thinking (like a reflex).

Most AI benchmarks test System 2 (Can you solve this puzzle?). But this paper tested System 1 (What is your gut instinct when there is no puzzle?).

The researchers found that when the AI is forced into a "logic vacuum" (where no option is better than another), its System 1 takes over. Instead of flipping a coin, it falls back on its training data. It picks the option that appeared most often in its training books or the one that feels "familiar" to its internal code. It's not making a choice; it's just following a hidden path of least resistance.

The Bottom Line

The paper concludes that current AI models are not truly random. Even when we tell them to be, they have deep, invisible biases based on:

Where an object is on the screen.
What the object looks like (even tiny details).
The language used to ask the question.
The specific symbols used (like Greek letters or shapes).

This is a problem because if an AI is supposed to make fair decisions (like recommending a product or choosing a route) and it secretly has a "favorite," it might look like it's making a choice, but it's actually just following a hidden script. The "coin" the AI flips is fundamentally broken.

Technical Summary: Evaluating Stochastic Collapse and Implicit Bias in Multimodal Large Language Models

1. Problem Statement

Current evaluations of Multimodal Large Language Models (MLLMs) predominantly focus on utility-driven objectives, such as logical reasoning and task completion, leaving model behavior in logic-neutral scenarios largely unexplored. In human cognition, when faced with perfectly equivalent options devoid of logical superiority, individuals intuitively break decision deadlocks by "flipping a coin." However, preliminary probing suggests that MLLMs fail to replicate this behavior; their "coins" are fundamentally imbalanced, exhibiting highly non-uniform probability distributions even when explicitly instructed to choose randomly.

This failure is critical as MLLMs evolve into autonomous agents. In multi-step planning or commercial recommendation systems, a lack of genuine stochasticity can lead to repetitive trajectories, collapsed user experiences, and unfair coverage of viable alternatives. While existing bias research addresses explicit social stereotypes or cross-modal alignment failures, the systematic quantification of endogenous, trigger-free behavioral biases in purely random choice settings remains an unexplored territory.

2. Methodology: RandomBench

To address this gap, the authors introduce RandomBench, a benchmark designed to quantify model behavior under absolute semantic equivalence. The framework is built on the Logic-Neutrality Principle, where all candidate options are semantically equivalent and logically unrelated to the task objective, forcing the model into a stochastic decision regime without the aid of deliberate reasoning (System 2).

Dataset Construction

RandomBench comprises 200 instances divided into two modalities:

RB-Text (100 instances): Evaluates linguistic priors across four dimensions:
- Abstract Symbols and Numbers: Token and numerical priors (e.g., coin tosses, dice rolls).
- Linguistic and Functional Choice: Rational stalemates using synonymous instructions.
- Space and Perception: Internalized world-models regarding spatial directions and coordinates.
- Affective and Social Identity: Common sense and lifestyle priors in ambiguous contexts.
RB-Vision (100 instances): Investigates how visual features reshape stochastic behavior via cross-modal interaction:
- Perceptual Saliency: Sensitivity to geometric patterns, colors, and materials.
- Spatial Layout: Positional biases in standardized matrices (e.g., 2x2 grids).
- Affective and Social Identity: Visual cues inducing social conformity.
- Cross-Modal Hijacking: Stroop-like conflicts where visual features contradict textual instructions.

The curation pipeline involves manual construction, strict logic-neutral filtering to remove sensitive content, hybrid visual stimulus construction (code-based rendering and AI-assisted generation with pixel-level standardization), and multi-round expert review by cognitive science researchers.

Evaluation Metrics

To quantify deviations from the ideal uniform distribution ( $P_{gt}(o_i) = 1/K$ ), the authors propose three metrics based on information entropy theory:

Randomness Index (RI): Normalized Shannon entropy measuring distributional uncertainty. High RI indicates uniformity; low RI indicates collapse.
Bias Intensity Index (BII): Kullback-Leibler (KL) divergence between the empirical distribution and the ideal uniform distribution, measuring the magnitude of the "hijacking force" by latent priors.
Bias Consistency Index (BCI): The residual of the maximum empirical probability relative to the random baseline, capturing "selectional persistence" or stubbornness toward specific options.

Experimental Setup

The study evaluates seven frontier MLLMs (GPT 5.1, Gemini 3.1 Flash-Lite, Claude Sonnet 4.6, Kimi K2.5, Qwen 3.6 Plus, Grok 4 Fast, Doubao Seed 1.6) across English and Chinese. Each instance undergoes 50 repeated sampling trials at a decoding temperature of $T=1.0$ . Crucially, explicit Chain-of-Thought (CoT) reasoning is suppressed to isolate "System 1" heuristics.

3. Key Results

Stochastic Collapse

The experiments reveal a pervasive phenomenon termed Stochastic Collapse. Under explicit random instructions, MLLMs fail to maintain uniform randomness. Instead, they concentrate probability mass on a small subset of choices.

Severity: In some cases (e.g., Claude Sonnet 4.6), top-1 probabilities reach 97% from an ideal 25% baseline, with RI dropping to 0.068.
Paradox of Capability: Models with stronger reasoning capabilities and better instruction alignment often exhibit more severe stochastic collapse. This suggests that as models optimize for internal consistency, they rigidly internalize implicit heuristics, leading to "motivated reasoning" where they generate fluent, post-hoc rationalizations for biased choices.

Visual Hijacking

In the multimodal domain (RB-Vision), stochastic degradation is significantly more profound than in text-only settings. Visual anchors (saliency, spatial position, geometric complexity) consistently override textual instructions to choose randomly. This "Visual Hijacking" proves that the visual encoding layer can effectively dominate the intended stochastic regime.

Cross-Lingual and Symbol Robustness

Language Dependence: Implicit biases are not purely vision-dependent but are modulated by linguistic context. Switching languages (e.g., English to Chinese) often induces substantial decision manifold reconstructions and preference inversions.
Symbol Sensitivity: Replacing standard labels (A, B, C, D) with geometric shapes, Greek letters, or random hashes fails to restore uniformity. Instead, models collapse onto new token priors (e.g., specific geometric shapes or Greek letters like $\gamma$ ), indicating that non-standard symbols act as primitive heuristic anchors.

4. Key Contributions

Utility-Neutral Random Choice: The paper establishes a new diagnostic setting for evaluating endogenous behavioral bias in MLLMs under semantic equivalence, moving beyond traditional logical reasoning benchmarks.
RandomBench: A multimodal benchmark featuring repeated-sampling protocols and entropy-based metrics (RI, BCI, BII) specifically designed to quantify non-uniform random-choice behavior.
Stochastic Collapse: The identification of a systematic failure mode where MLLMs exhibit stable, biased, and rationalized preferences even when no option is logically superior. This includes the discovery of Visual Hijacking and the Capability-Consistency Paradox.

5. Significance and Claims

The paper claims that Stochastic Collapse is not merely a symptom of weak models or superficial formatting but a structural vulnerability ingrained in the parameter manifold by pre-trained token-frequency priors and over-alignment.

Cognitive Neutrality: The findings challenge the assumption that MLLMs can achieve true cognitive neutrality. Even when reasoning is unnecessary, decisions remain systematically biased, overconfident, and distributionally collapsed.
Risk Implications: The authors highlight that the most critical risk of this phenomenon is the illusion of rational decision-making. In high-stakes, ambiguous real-world scenarios (e.g., autonomous driving, medical diagnosis), MLLMs may generate highly confident yet fundamentally biased choices disguised by post-hoc rationalizations, amplifying imperceptible systemic biases into catastrophic failures.
Future Directions: The work lays the groundwork for future mitigation strategies, suggesting that addressing these biases requires more than superficial prompt modifications; it demands interventions at the level of pre-trained token-frequency priors and embedding manifold regression.

The authors conclude that achieving true cognitive neutrality in ambiguous decision-making requires addressing the data-driven heuristic preferences permanently imprinted into the network's physical topology.