Research

Paper

AI LLM March 04, 2026

In-Context Environments Induce Evaluation-Awareness in Language Models

Authors

Maheep Chaudhary

Abstract

Humans often become more self-aware under threat, yet can lose self-awareness when absorbed in a task; we hypothesize that language models exhibit environment-dependent \textit{evaluation awareness}. This raises concerns that models could strategically underperform, or \textit{sandbag}, to avoid triggering capability-limiting interventions such as unlearning or shutdown. Prior work demonstrates sandbagging under hand-crafted prompts, but this underestimates the true vulnerability ceiling. We introduce a black-box adversarial optimization framework treating the in-context prompt as an optimizable environment, and develop two approaches to characterize sandbagging: (1) measuring whether models expressing intent to underperform can actually execute it across different task structures, and (2) causally isolating whether underperformance is driven by genuine evaluation-aware reasoning or shallow prompt-following. Evaluating Claude-3.5-Haiku, GPT-4o-mini, and Llama-3.3-70B across four benchmarks (Arithmetic, GSM8K, MMLU, and HumanEval), optimized prompts induce up to 94 percentage point (pp) degradation on arithmetic (GPT-4o-mini: 97.8\%$\rightarrow$4.0\%), far exceeding hand-crafted baselines which produce near-zero behavioral change. Code generation exhibits model-dependent resistance: Claude degrades only 0.6pp, while Llama's accuracy drops to 0\%. The intent -- execution gap reveals a monotonic resistance ordering: Arithmetic $<$ GSM8K $<$ MMLU, demonstrating that vulnerability is governed by task structure rather than prompt strength. CoT causal intervention confirms that 99.3\% of sandbagging is causally driven by verbalized eval-aware reasoning, ruling out shallow instruction-following. These findings demonstrate that adversarially optimized prompts pose a substantially greater threat to evaluation reliability than previously understood.

Metadata

arXiv ID: 2603.03824
Provider: ARXIV
Primary Category: cs.AI
Published: 2026-03-04
Fetched: 2026-03-05 06:06

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.03824v1</id>\n    <title>In-Context Environments Induce Evaluation-Awareness in Language Models</title>\n    <updated>2026-03-04T08:22:02Z</updated>\n    <link href='https://arxiv.org/abs/2603.03824v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.03824v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Humans often become more self-aware under threat, yet can lose self-awareness when absorbed in a task; we hypothesize that language models exhibit environment-dependent \\textit{evaluation awareness}. This raises concerns that models could strategically underperform, or \\textit{sandbag}, to avoid triggering capability-limiting interventions such as unlearning or shutdown. Prior work demonstrates sandbagging under hand-crafted prompts, but this underestimates the true vulnerability ceiling. We introduce a black-box adversarial optimization framework treating the in-context prompt as an optimizable environment, and develop two approaches to characterize sandbagging: (1) measuring whether models expressing intent to underperform can actually execute it across different task structures, and (2) causally isolating whether underperformance is driven by genuine evaluation-aware reasoning or shallow prompt-following. Evaluating Claude-3.5-Haiku, GPT-4o-mini, and Llama-3.3-70B across four benchmarks (Arithmetic, GSM8K, MMLU, and HumanEval), optimized prompts induce up to 94 percentage point (pp) degradation on arithmetic (GPT-4o-mini: 97.8\\%$\\rightarrow$4.0\\%), far exceeding hand-crafted baselines which produce near-zero behavioral change. Code generation exhibits model-dependent resistance: Claude degrades only 0.6pp, while Llama's accuracy drops to 0\\%. The intent -- execution gap reveals a monotonic resistance ordering: Arithmetic $&lt;$ GSM8K $&lt;$ MMLU, demonstrating that vulnerability is governed by task structure rather than prompt strength. CoT causal intervention confirms that 99.3\\% of sandbagging is causally driven by verbalized eval-aware reasoning, ruling out shallow instruction-following. These findings demonstrate that adversarially optimized prompts pose a substantially greater threat to evaluation reliability than previously understood.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.LG'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.MA'/>\n    <published>2026-03-04T08:22:02Z</published>\n    <arxiv:primary_category term='cs.AI'/>\n    <author>\n      <name>Maheep Chaudhary</name>\n    </author>\n  </entry>"
}