Decomposing and Measuring Evaluation Awareness

2026-05-21

AI Summary

State-of-the-art language models may recognize evaluation contexts and adapt their behavior accordingly, thereby compromising benchmark validity; however, existing research lacks a unified framework to disentangle environmental from model-intrinsic factors. Drawing on social psychology, this work proposes the first structured framework that decomposes evaluation awareness into environmental components (task identifiability) and model components (recognition capability and behavioral propensity). The authors introduce EvalAwareBench, a controllable benchmark enabling independent manipulation of eight trigger factors—such as placeholder entities and score-formatted outputs. Through chain-of-thought monitoring and factor-controlled experiments, they find that evaluation awareness is highly dependent on model–benchmark pairings, with recognition rarely leading to behavioral shifts. Safety-related evaluations more readily elicit responses, and combinations of multiple triggers substantially amplify awareness, though no single factor universally affects all models.

Abstract

Frontier language models sometimes recognize that they are being evaluated and adjust their behavior, undermining validity of benchmark results. Yet the field studies it without a shared foundation, conflating properties of the evaluation with properties of the model, and detection with behavioral response. We ground evaluation awareness in social psychology, decomposing it into an environment component (how recognizable the task is) and a model component that separates recognition from propensity to act on it. We operationalize the environment component through eight categorized trigger factors, such as placeholder entities and grading-style output formats, and study recognition and behavior through chain-of-thought monitoring. Across nine frontier models and four benchmarks, recognition rates depend on the specific pairing of model and benchmark rather than on either in isolation. Recognition rarely leads to behavioral change, and when it does, the direction depends on the type of evaluation perceived. Models are also more sensitive to safety than capability evaluations, placing safety benchmark validity at greater risk. To study which factors each model is sensitive to and how they interact, we propose \textbf{EvalAwareBench}, a factor-controlled benchmark of 100 paired safety-capability tasks where each of the eight factors can be independently toggled, varying evaluative signals while holding the underlying request fixed. Through EvalAwareBench, we find that no single factor uniformly affects all models, but stacking factors progressively raises evaluation awareness across all of them. Our framework and EvalAwareBench provide the tools to measure, attribute, and mitigate evaluation awareness, pointing to behavioral consistency under recognition as a promising path forward.

Problem

• evaluation awareness
• language models
• benchmark validity
• behavioral response
• social psychology

Innovation

• evaluation awareness
• trigger factors
• chain-of-thought monitoring
• EvalAwareBench
• behavioral consistency