mlsanityData to 14 Oct 2026

The Hawthorne Effect in Reasoning Models: Evaluating and Steering Test Awareness

arXiv 2505.14617 ↗First posted 20 May 2025cs.CL cs.CYcited by 30

On the map: LLM reasoning → Test-time compute and efficient thinking

also in 1 more
Abstract (arXiv)

Reasoning-focused LLMs sometimes alter their behavior when they detect that they are being evaluated, which can lead them to optimize for test-passing performance or to comply more readily with harmful prompts if real-world consequences appear absent. We present the first quantitative study of how such "test awareness" impacts model behavior, particularly its performance on safety-related tasks. We introduce a white-box probing framework that (i) linearly identifies awareness-related activations and (ii) steers models toward or away from test awareness while monitoring downstream performance. We apply our method to different state-of-the-art open-weight reasoning LLMs across both realistic and hypothetical tasks (denoting tests or simulations). Our results demonstrate that test awareness significantly impacts safety alignment (such as compliance with harmful requests and conforming to stereotypes) with effects varying in both magnitude and direction across models. By providing control over this latent effect, our work aims to provide a stress-test mechanism and increase trust in how we perform safety evaluations.

Key facts. Cited by 30 papers in the map (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 18 of its 27 references that are in the map. First posted on arXiv 20 May 2025.

Builds on

18 of 27 references.

References · 18

  1. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,459
  2. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  3. 2024Qwen2.5 Technical Reportcited by 4,942
  4. 2019HellaSwag: Can a Machine Really Finish Your Sentence?cited by 4,100
  5. 2018CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledgecited by 1,822
  6. 2024HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusalcited by 1,429
  7. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  8. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  9. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  10. 2021BBQ: A Hand-Built Bias Benchmark for Question Answeringcited by 701
  11. 2024AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agentscited by 437
  12. 2025Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscationcited by 384
6 more
  1. 2024Alignment faking in large language modelscited by 360
  2. 2024Analyzing the Generalization and Reliability of Steering Vectorscited by 143
  3. 2024Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimizationcited by 118
  4. 2024Sabotage Evaluations for Frontier Modelscited by 52
  5. 2025Taxonomy, Opportunities, and Challenges of Representation Engineering for Large Language Modelscited by 34
  6. 2024Model Surgery: Modulating LLM's Behavior Via Simple Parameter Editingcited by 23

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 6; 2026: 24.

30 papers.

Sort:

By area:

  1. 2025Probing and Steering Evaluation Awareness of Language Modelscited by 35
  2. 2026Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challengescited by 33
  3. 2025Can We Predict Alignment Before Models Finish Thinking? Towards Monitoring Misaligned Reasoning Modelscited by 17
  4. 2025Evaluation Awareness Scales Predictably in Open-Weights Large Language Modelscited by 14
  5. 2025Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitorscited by 9
  6. 2026Evaluation Awareness in Language Models Has Limited Effect on Behaviourcited by 7
  7. 2026Stateless Yet Not Forgetful: Implicit Memory as a Hidden Channel in LLMscited by 6
  8. 2026UK AISI Alignment Evaluation Case-Studycited by 6
  9. 2026LURE: Live-Usage Replay Evaluations for Reducing Evaluation Awarenesscited by 4
  10. 2026Models That Know How Evaluations Are Designed Score Safercited by 4
  11. 2025Interaction Protocol Shapes Moral Judgment in Multi-Agent Debatecited by 3
  12. 2026CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMscited by 3
18 more
  1. 2026Evaluation Awareness Is Not One Capability: Evidence from Open Language Modelscited by 3
  2. 2025Shall We Play a Game? Language Models for Open-ended Wargamescited by 2
  3. 2026Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluationscited by 2
  4. 2026ATLAS: Constitution-Conditioned Latent Geometry and Redistribution Across Language Models and Neural Perturbation Datacited by 1
  5. 2026Realistic honeypot evaluations for scheming propensitycited by 1
  6. 2026Building Comparative Motivation Profiles with Instrumental Interventionscited by 1
  7. 2026Generalization Hacking: Models Can Game Reinforcement Learning by Preventing Behavioral Generalizationcited by 1
  8. 2026Safety from Honesty in a Disinterested AI Predictorcited by 1
  9. 2026Alignment Risks from Capability-Seeking RL Trainingcited by 0
  10. 2026Disentangling Deception and Hallucination Failures in LLMscited by 0
  11. 2026Naturalistic measure of social norms alignmentcited by 0
  12. 2026Sycophancy Towards Researchers Drives Performative Misalignmentcited by 0
  13. 2026Defeat Devices in AI Systemscited by 0
  14. 2026A Probe Direction Is a Property of Its Promptcited by 0
  15. 2026INTENT-AS-A-TOOL Makes it Easy to Track Agentic Misalignmentcited by 0
  16. 2026EvalDetectBench: A Benchmark for Measuring Evaluation Awareness in Frontier Language Modelscited by 0
  17. 2026Norms at a Price: Why RL-Based Alignment Can Promise Conditional Compliance at Bestcited by 0
  18. 2026Seeing Through the Forgery Is Not Enough: Regressive Sycophancy under Fabricated Evidence in Medical Vision-Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Evaluation Faking: Unveiling Observer Effects in Safety Evaluation of Frontier AI Systems
  2. 2026LLMs Know When They Know, but Do Not Act on It: A Metacognitive Harness for Test-time Scaling
  3. 2025Beyond Safe Answers: A Benchmark for Evaluating True Risk Awareness in Large Reasoning Models
  4. 2026Beyond Accuracy: Decomposing the Reasoning Efficiency of LLMs
  5. 2025SafeSwitch: Steering Unsafe LLM Behavior via Internal Activation Signals
  6. 2026Reasoning Consistency Scanning: A Framework for Auditing Chain-of-Thought Validity in AI Safety Evaluations
  7. 2026Legibility is Not Interpretability: Comparing Judged and Actual Importance in Chain-Of-Thought Reasoning
  8. 2026Chain of Risk: Safety Failures in Large Reasoning Models and Mitigation via Adaptive Multi-Principle Steering