In-Context Environments Induce Evaluation-Awareness in Language Models

• URL: http://arxiv.org/abs/2603.03824v1
• Date: Wed, 04 Mar 2026 08:22:02 GMT
• Title: In-Context Environments Induce Evaluation-Awareness in Language Models
• Authors: Maheep Chaudhary,
• Abstract summary: Humans often become more self-aware under threat, yet can lose self-awareness when absorbed in a task.<n>We introduce a black-box adversarial optimization framework treating the in-context prompt as an optimizable environment.<n>We show that adversarially optimized prompts pose a substantially greater threat to evaluation reliability than previously understood.
• Score: 0.12691047660244334
• License: http://creativecommons.org/licenses/by/4.0/
• 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.

Related papers

• Reinforcement Inference: Leveraging Uncertainty for Self-Correcting Language Model Reasoning [0.0]
  Reinforcement Inference uses the model's own uncertainty to selectively invoke a second, more deliberate reasoning attempt.<n>On 12,032 MMLU-Pro questions across 14 subjects, using DeepSeek-v3.2 with deterministic decoding in a zero-shot setting, Reinforcement Inference improves accuracy from 60.72% to 84.03%.
  arXiv  Detail & Related papers  (2026-02-09T11:08:24Z)
• RoguePrompt: Dual-Layer Ciphering for Self-Reconstruction to Circumvent LLM Moderation [0.0]
  This paper presents an automated jailbreak attack that converts a disallowed user query into a self reconstructing prompt.<n>We instantiate RoguePrompt against GPT 4o and evaluate it on 2 448 prompts that a production moderation system previously marked as strongly rejected.<n>Under an evaluation protocol that separates three security relevant outcomes bypass, reconstruction, and execution the attack attains 84.7 percent bypass, 80.2 percent reconstruction, and 71.5 percent full execution.
  arXiv  Detail & Related papers  (2025-11-24T05:42:54Z)
• Parrot: Persuasion and Agreement Robustness Rating of Output Truth -- A Sycophancy Robustness Benchmark for LLMs [0.0]
  PARROT (Persuasion and Agreement Robustness Rating of Output Truth) is a robustness focused framework designed to measure the degradation in accuracy under social pressure exerted on users.<n>We evaluate 22 models using 1,302 MMLU-style multiple-choice questions across 13 domains and domain-specific authority templates.
  arXiv  Detail & Related papers  (2025-11-21T13:01:28Z)
• DeLeaker: Dynamic Inference-Time Reweighting For Semantic Leakage Mitigation in Text-to-Image Models [55.30555646945055]
  Text-to-Image (T2I) models are vulnerable to semantic leakage.<n>We introduce DeLeaker, a lightweight approach that mitigates leakage by directly intervening on the model's attention maps.<n>SLIM is the first dataset dedicated to semantic leakage.
  arXiv  Detail & Related papers  (2025-10-16T17:39:21Z)
• Think Twice, Generate Once: Safeguarding by Progressive Self-Reflection [18.467741067831877]
  We introduce Progressive Self-Reflection, a novel inference-time technique that empowers large language models to self-monitor and correct their outputs dynamically.<n> Experimental results demonstrate that applying our proposed method to Llama-3.1-8B-Instruct reduces the attack success rate from 77.5% to 5.9%.<n>Our approach acts as a test-time scaling method, where additional self-reflection rounds enhance safety at the cost of inference overhead.
  arXiv  Detail & Related papers  (2025-09-29T12:54:28Z)
• DiffuGuard: How Intrinsic Safety is Lost and Found in Diffusion Large Language Models [50.21378052667732]
  We conduct an in-depth analysis of dLLM vulnerabilities to jailbreak attacks across two distinct dimensions: intra-step and inter-step dynamics.<n>We propose DiffuGuard, a training-free defense framework that addresses vulnerabilities through a dual-stage approach.
  arXiv  Detail & Related papers  (2025-09-29T05:17:10Z)
• From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs [58.02809208460186]
  We revisit this paradox using high-quality traces from DeepSeek-R1 as demonstrations.<n>We find that adding more exemplars consistently degrades accuracy, even when demonstrations are optimal.<n>We introduce Insight-to-solve (I2S), a sequential test-time procedure that turns demonstrations into explicit, reusable insights.
  arXiv  Detail & Related papers  (2025-09-27T08:59:31Z)
• Cognitive Load Limits in Large Language Models: Benchmarking Multi-Hop Reasoning [0.0]
  Large Language Models (LLMs) excel at isolated tasks, but their reasoning under cognitive load remains poorly understood.<n>We introduce a formal theory of computational cognitive load, positing that extraneous, task-irrelevant information (Context Saturation) and interference from task-switching are key mechanisms that degrade performance.
  arXiv  Detail & Related papers  (2025-09-23T19:36:56Z)
• Reasoning Models Are More Easily Gaslighted Than You Think [85.84943447589511]
  We evaluate three state-of-the-art reasoning models, including OpenAI's o4-mini, Claude-3.7-Sonnet and Gemini-2.5-Flash.<n>Our evaluation reveals significant accuracy drops following gaslighting negation prompts.<n>We introduce GaslightingBench-R, a new diagnostic benchmark designed to evaluate reasoning models' susceptibility to defend their belief.
  arXiv  Detail & Related papers  (2025-06-11T12:52:25Z)
• Benchmarking Reasoning Robustness in Large Language Models [76.79744000300363]
  We find significant performance degradation on novel or incomplete data.<n>These findings highlight the reliance on recall over rigorous logical inference.<n>This paper introduces a novel benchmark, termed as Math-RoB, that exploits hallucinations triggered by missing information to expose reasoning gaps.
  arXiv  Detail & Related papers  (2025-03-06T15:36:06Z)
• Turning Logic Against Itself : Probing Model Defenses Through Contrastive Questions [50.40122190627256]
  We introduce POATE, a novel jailbreak technique that harnesses contrastive reasoning to provoke unethical responses.<n>PoATE crafts semantically opposing intents and integrates them with adversarial templates, steering models toward harmful outputs with remarkable subtlety.<n>To counter this, we propose Intent-Aware CoT and Reverse Thinking CoT, which decompose queries to detect malicious intent and reason in reverse to evaluate and reject harmful responses.
  arXiv  Detail & Related papers  (2025-01-03T15:40:03Z)
• On the Worst Prompt Performance of Large Language Models [93.13542053835542]
  Performance of large language models (LLMs) is acutely sensitive to the phrasing of prompts. We introduce RobustAlpacaEval, a new benchmark that consists of semantically equivalent case-level queries. Experiments on RobustAlpacaEval with ChatGPT and six open-source LLMs from the Llama, Mistral, and Gemma families uncover substantial variability in model performance.
  arXiv  Detail & Related papers  (2024-06-08T13:40:38Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.