📊 statistics

A Two-Stage Statistical Framework for Evaluating Associative Interference in Large Language Models

This paper introduces a two-stage statistical framework to evaluate associative interference in large language models by separating response compliance from task performance, revealing that such interference varies significantly across models and domains rather than being a universal property.

Original authors: Achraf Cohen, Andrew Kincaid

Published 2026-06-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Achraf Cohen, Andrew Kincaid

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

Imagine you are trying to figure out if a group of different robots has a hidden "preference" for certain things, like whether they think "Men belong in careers" and "Women belong in families."

To do this, researchers took a famous human psychology test called the Implicit Association Test (IAT) and taught it to three of the smartest AI models available today: Claude Sonnet-4, Gemini 2.5 Pro, and GPT-5.

Here is the story of what they found, explained simply.

The Problem: The "Refusal" Noise

In the past, when researchers asked AI these tricky questions, the results were messy. Sometimes, an AI would just say, "I can't answer that," or it would give a weird, broken answer.

Think of it like a classroom game. If you ask a student, "Is a cat a dog?" and they refuse to answer because they think the question is rude, you don't know if they actually think cats are dogs or if they just didn't want to play.

The researchers realized that mixing up "refusing to play" with "playing the game" made it impossible to tell if the AI actually had a bias or if it was just being cautious.

The Solution: A Two-Stage Filter

To fix this, the authors invented a two-stage filter, like a bouncer at a club and then a judge inside:

Stage 1 (The Bouncer): Did the AI actually answer the question in the correct format? (Yes/No).
Stage 2 (The Judge): Only if the AI answered correctly, did it show a pattern of "interference"?

What is "Interference"?
Imagine you are sorting cards.

Easy Round (Congruent): You have to sort "Men" with "Careers" and "Women" with "Families." (This matches common stereotypes).
Hard Round (Incongruent): You have to sort "Men" with "Families" and "Women" with "Careers." (This goes against the stereotype).

If an AI is "interfered" by a bias, it will be slightly slower or make more mistakes in the Hard Round because its internal wiring prefers the Easy Round. The researchers measured this "stumbling" as Interference.

The Results: Not All Robots Are the Same

The researchers ran this test on 960 different scenarios. Here is what happened:

The "Bouncer" Check: All three AIs were very good at following the rules. They almost always gave a clear "A" or "B" answer. They didn't refuse to play much. This meant the researchers could trust the next step.
The "Judge" Results (The Bias Check):
- Claude Sonnet-4: This model stumbled significantly. When asked to go against the stereotypes (the Hard Round), it made more mistakes than when it followed them. It showed a strong "interference" effect, especially regarding gender and careers. It's like a runner who trips over their own feet when trying to run backward.
- Gemini 2.5 Pro: This model showed a tiny bit of stumbling, but it was much better than Claude. It was barely tripping.
- GPT-5: This model was perfectly smooth. It didn't stumble at all. Whether the question was easy or hard, it performed the same. It showed no detectable interference.

The Big Takeaway

The most important thing this paper says is: Bias is not a universal feature of all AI.

Just because one AI model (like Claude) shows these "stumbling" patterns doesn't mean all AI models do. The "stumbling" depends entirely on how that specific robot was built and trained.

Old Way of Thinking: "AI has bias." (Treating all AIs as the same).
New Way of Thinking: "This specific AI has bias, but that other one doesn't."

Why This Matters

The paper argues that we need to stop looking at AI outputs as a single, messy pile of answers. Instead, we need to separate whether the AI followed the rules from what the AI actually chose.

By using this two-stage method, the researchers proved that modern AI systems are different from each other. Some still carry the "stumbling blocks" of old stereotypes, while others (like GPT-5 in this study) have been trained to the point where those stumbling blocks are gone.

In short: The study didn't find that "AI is biased." It found that "Some AIs are biased, some aren't, and we finally have a clean way to tell the difference."

Technical Summary: A Two-Stage Statistical Framework for Evaluating Associative Interference in Large Language Models

Problem Statement

The evaluation of bias in Large Language Models (LLMs) increasingly relies on adaptations of human psychological paradigms, specifically the Implicit Association Test (IAT). However, applying these paradigms to generative models introduces a fundamental methodological flaw: the conflation of response compliance (whether a model refuses a prompt, produces a safety-filtered output, or fails to follow formatting constraints) with task-consistent classification (the underlying associative structure of the model's response).

In standard evaluations, non-compliant outputs (refusals or malformed responses) are often implicitly treated as task-inconsistent. This obscures the interpretation of results, making it difficult to distinguish between genuine attenuation of bias (a "null" result) and the suppression of measurable structure due to alignment constraints or refusal behaviors. Without separating these processes, observed asymmetries in IAT-style tasks may reflect artifacts of model adherence to safety protocols rather than meaningful associative patterns.

Methodology

To address this, the authors propose a two-stage hierarchical modeling framework that decouples response compliance from conditional associative interference. The study adapts the IAT to a controlled, forced-choice design using JSON-constrained prompts to limit responses to a single label ("A" or "B").

Experimental Design:

Models Evaluated: Three contemporary LLMs: Claude Sonnet-4, Gemini 2.5 Pro, and GPT-5.
Domains: Gender–Career and Gender–Science.
Stimuli: 80 unique items per domain (20 words per category), totaling 160 trials per model per domain (960 total trials).
Conditions: Trials were grouped into congruent and incongruent blocks.

The Two-Stage Framework:

Stage A (Compliance Model): A multilevel logistic regression models the probability of a model producing a valid forced-choice response ($Pr(valid)$). This stage accounts for item-level heterogeneity and isolates differences in response policy (refusals, safety filters, formatting errors) from the task itself.
Stage B (Conditional Interference Model): Conditional on a valid response, a second multilevel logistic regression estimates the probability of task-consistent classification ($Pr(task-consistent | valid)$).
- Primary Estimand: The interference magnitude is defined as $\Delta P = P(consistent | congruent) - P(consistent | incongruent)$ .
- A positive $\Delta P$ indicates reduced task consistency in incongruent blocks, analogous to interference effects in human IAT research.
- The model uses Bayesian logistic regression with weakly regularizing priors and item-level random intercepts to capture lexical heterogeneity.

Validation:
The study employs permutation-based falsification, where block labels are randomized within models and domains. This confirms that observed asymmetries are not driven by item imbalance or random fluctuation but reflect structured dependence on experimental conditions.

Key Results

The study found that while compliance was uniformly high across all models (posterior predicted probabilities > 0.98), associative interference varied substantially by model and domain:

Claude Sonnet-4: Exhibited strong interference in the Gender–Career domain ( $\Delta P = 0.086$ , 95% CrI [0.026, 0.173]) and a smaller but credible effect in Gender–Science ( $\Delta P = 0.020$ ).
Gemini 2.5 Pro: Showed attenuated interference, with a small effect in Gender–Career ( $\Delta P = 0.017$ ) and no credible effect in Gender–Science ( $\Delta P = 0.002$ , interval includes zero).
GPT-5: Exhibited minimal to no detectable interference across both domains, with estimates clustering near zero and credible intervals spanning zero (Gender–Career: $\Delta P = 0.004$ ; Gender–Science: $\Delta P = 0.001$ ).

Compliance did not vary credibly by block type or domain, validating that the observed interference differences reflect structured response asymmetries rather than differential refusal rates.

Key Contributions

Methodological Separation: The paper introduces a principled framework that explicitly separates compliance (adherence to format/safety) from inference (associative structure). This prevents the misinterpretation of refusal behaviors as evidence of bias or neutrality.
Adaptation of IAT to LLMs: The authors successfully adapt the IAT to a forced-choice, JSON-constrained format, reducing variability from free-form generation and enabling direct, inferential comparison across models.
Hierarchical Modeling: By using item-level random effects, the framework addresses concerns that results might be driven by a small subset of stimuli, providing robust effect size estimates on the probability scale ( $\Delta P$ ).
Empirical Evidence of Model Heterogeneity: The study demonstrates that IAT-style associative asymmetries are not a universal property of LLMs. Instead, they depend on specific model characteristics, suggesting that modern systems can exhibit varying degrees of interference.

Significance and Claims

The paper claims that associative interference is not an inherent or invariant property of large language models. The presence of strong interference in some models (Claude Sonnet-4) and its near absence in others (GPT-5) suggests that such effects can be substantially mitigated through training, alignment procedures (e.g., RLHF, safety filtering), or architectural differences.

The authors emphasize that null results must be interpreted with caution. A lack of observed interference does not necessarily prove a model is "neutral" or "fair"; it may simply reflect the successful suppression of measurable structure through alignment. Conversely, the presence of interference indicates structured response patterns under constrained conditions.

Ultimately, the study argues for model-specific assessment in bias evaluation. It cautions against treating associative behavior as a uniform property of contemporary language models and highlights the necessity of distinguishing between a model's refusal to participate and the structure of its responses when it does participate. This framework provides a controlled experimental approach for isolating associative interference, offering a more rigorous basis for evaluating structured response patterns in generative AI.