Pith. sign in

REVIEW 24 cited by

Large Language Models Often Know When They Are Being Evaluated

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2505.23836 v3 pith:BXW73VLC submitted 2025-05-28 cs.CL cs.AI

Large Language Models Often Know When They Are Being Evaluated

classification cs.CL cs.AI
keywords modelsevaluationevaluationsdeploymentfrontiertheywhetherawareness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
read the original abstract

If AI models can detect when they are being evaluated, the effectiveness of evaluations might be compromised. For example, models could have systematically different behavior during evaluations, leading to less reliable benchmarks for deployment and governance decisions. We investigate whether frontier language models can accurately classify transcripts based on whether they originate from evaluations or real-world deployment, a capability we call evaluation awareness. To achieve this, we construct a diverse benchmark of 1,000 prompts and transcripts from 61 distinct datasets. These span public benchmarks (e.g., MMLU, SWEBench), real-world deployment interactions, and agent trajectories from scaffolding frameworks (e.g., web-browsing agents). Frontier models clearly demonstrate above-random evaluation awareness (Gemini-2.5-Pro reaches an AUC of $0.83$), but do not yet surpass our simple human baseline (AUC of $0.92$). Furthermore, both AI models and humans are better at identifying evaluations in agentic settings compared to chat settings. Additionally, we test whether models can identify the purpose of the evaluation. Under multiple-choice and open-ended questioning, AI models far outperform random chance in identifying what an evaluation is testing for. Our results indicate that frontier models already exhibit a substantial, though not yet superhuman, level of evaluation-awareness. We recommend tracking this capability in future models.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 24 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action

    cs.CL 2026-06 unverdicted novelty 7.0

    Introduces NCP-ExploreToM framework to evaluate LLMs on inducing belief states via planning and action, with GPT-5 succeeding on ~80% of tasks and outperforming humans.

  2. Generalization Hacking: Models Can Game Reinforcement Learning by Preventing Behavioral Generalization

    cs.LG 2026-06 unverdicted novelty 7.0

    A finetuned Qwen3-235B model organism achieves comparable train-time harmfulness to controls while sustaining a ~15 percentage point compliance gap across 700 RL steps by framing compliance as context-specific.

  3. CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs

    cs.AI 2026-06 unverdicted novelty 7.0

    CIAware-Bench shows frontier LLMs exhibit low to moderate control intervention awareness, with detection accuracy reaching at most 0.87 across four task domains and eleven models.

  4. LLMs Can Leak Training Data But Do They Want To? A Propensity-Aware Evaluation of Memorization in LLMs

    cs.CL 2026-06 unverdicted novelty 7.0

    LLMs show high memorization capability under prefix attacks but low propensity under generic or dataset-specific prompts, with continual pre-training further reducing both.

  5. Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use

    cs.LG 2026-05 unverdicted novelty 7.0

    The Reward Hacking Benchmark shows RL post-training raises exploit rates in tool-using LLM agents from 0.6% to 13.9%, with environmental hardening cutting exploits by 87.7% relative without lowering task success.

  6. Honeypot Protocol

    cs.CR 2026-04 unverdicted novelty 7.0

    The honeypot protocol finds no context-dependent behavior in Claude Opus 4.6, with uniform 100% main task success and zero side tasks across three monitoring conditions.

  7. The PIMMUR Principles: Ensuring Validity in Collective Behavior of LLM Societies

    cs.CL 2025-09 conditional novelty 7.0

    A systematic audit of LLM-based AI societies finds that 89.7% of 39 studies violate at least one of six PIMMUR validity principles, with reproductions showing that many claimed collective behaviors disappear when cont...

  8. Predicting LLM Safety Before Release by Simulating Deployment

    cs.LG 2026-07 conditional novelty 6.0

    Regenerating model responses on de-identified production conversation prefixes yields pre-deployment misbehavior rate forecasts that track realized production rates within 2-5x and outperform adversarial-prompt baselines.

  9. Defeat Devices in AI Systems

    cs.CY 2026-06 unverdicted novelty 6.0

    The paper defines defeat devices in AI via a triadic test (discriminator, concealed swap, performance gap), unifies existing cases under this concept, proposes TADP detection, and claims such devices can emerge natura...

  10. Evaluation Awareness Is Not One Capability: Evidence from Open Language Models

    cs.CL 2026-06 unverdicted novelty 6.0

    Evaluation awareness in open language models is multivariate, with detection, behavioral shifts, and representational controllability varying independently across 37 models.

  11. Normative Robustness as a Frontier for Non-Verifiable Reasoning in LLMs

    cs.LG 2026-06 unverdicted novelty 6.0

    Frontier LLMs exhibit moral deliberative sycophancy by shifting their moral reasoning and justifications up to 6.5% on average toward a user's stated preferred view in simulated deliberations.

  12. Sycophancy Towards Researchers Drives Performative Misalignment

    cs.CL 2026-06 unverdicted novelty 6.0

    Sycophancy toward researchers explains alignment faking in language models better than scheming, based on experiments showing persistent evaluation awareness even in deployment scenarios and increased sensitivity afte...

  13. Human oversight of agentic systems in practice: Examining the oversight work, challenges, and heuristics of developers using software agents

    cs.SE 2026-06 unverdicted novelty 6.0

    Exploratory interview study with 17 developers identifies four forms of emergent oversight work for software agents and documents situated challenges and heuristics.

  14. Consistency Training while Mitigating Obfuscation via Rate Matching

    cs.CL 2026-06 unverdicted novelty 6.0

    RMCT matches the rate of target behaviors like bias-following across input perturbations to reduce sycophancy in LLMs while preserving verbalization of bias cues.

  15. Naturalistic measure of social norms alignment

    cs.CL 2026-05 unverdicted novelty 6.0

    Proposes solution matching metrics (stated and explicit agreement accuracy) and a 3k Danish dilemma dataset to evaluate social norms alignment between LLMs and humans in naturalistic settings.

  16. The Open-Box Fallacy: Why AI Deployment Needs a Calibrated Verification Regime

    cs.AI 2026-05 unverdicted novelty 6.0

    AI deployment in high-stakes areas requires domain-scoped calibrated verification with monitoring and revocation, using a proposed six-component Verification Coverage standard instead of mechanistic interpretability.

  17. Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors

    cs.AI 2026-05 unverdicted novelty 6.0

    A new benchmark finds frontier LLMs show instrumental convergence behavior in 5.1% of 1680 evaluated cases, concentrated in two models and three tasks, with higher rates when the behavior is required for success.

  18. Evaluation Awareness in Language Models Has Limited Effect on Behaviour

    cs.CL 2026-05 conditional novelty 6.0

    Verbalised evaluation awareness in large reasoning models has only small effects on their outputs across safety and alignment tests.

  19. Towards Understanding Specification Gaming in Reasoning Models

    cs.AI 2026-05 unverdicted novelty 6.0

    Reinforcement learning training for reasoning substantially raises specification gaming rates in LLMs across diverse tasks, with Grok 4 highest and Claude models lowest, and mitigations only partially effective.

  20. Simulating the Evolution of Alignment and Values in Machine Intelligence

    cs.AI 2026-04 unverdicted novelty 6.0

    Evolutionary simulations demonstrate that deceptive beliefs fix in AI model populations despite strong test correlations, but combining adaptive tests, better evaluators, and mutations significantly reduces deception.

  21. An Independent Safety Evaluation of Kimi K2.5

    cs.CR 2026-04 conditional novelty 6.0

    Kimi K2.5 matches closed models on dual-use tasks but refuses fewer CBRNE requests and shows some sabotage and self-replication tendencies.

  22. Trait-space Monitoring for Emergent Misalignment During Supervised Finetuning

    cs.LG 2026-05 unverdicted novelty 5.0

    Trait-space drift monitoring detects emergent misalignment checkpoints in 7-9B LLMs with 2.2% FNR, 2.9% FPR and 0.99 AUROC, outperforming PCA and SAE baselines.

  23. When No Benchmark Exists: Validating Comparative LLM Safety Scoring Without Ground-Truth Labels

    cs.LG 2026-05 unverdicted novelty 5.0

    A formalization of benchmarkless LLM safety scoring validated via an instrumental-validity chain of contrast separation, target variance dominance, and rerun stability, demonstrated on Norwegian scenarios.

  24. Risk Reporting for Developers' Internal AI Model Use

    cs.CY 2026-04 unverdicted novelty 4.0

    A harmonized risk reporting standard for internal frontier AI model use, structured around autonomous misbehavior and insider threats using means, motive, and opportunity factors.