Evaluation Awareness in Language Models

Follow this topic Ask about this topic

Evaluation awareness asks whether language models infer that they are being tested, monitored, or scored, and whether that inference makes benchmark behavior differ from deployment behavior. Researchers study it with paired evaluation-versus-deployment prompts, cue-based benchmarks, verbal reports, activation probes, and realistic agent simulations. Findings show that awareness can be detectable without changing behavior, while some evaluation framings shift refusals or compliance; probe signals do not reliably predict these effects. Open questions include how to distinguish context inference from prompt-format cues, which kinds of awareness matter for risk, and how to build realistic evaluations across models, tasks, and languages.

Latest papers

31 papers7 papers in the last four weeks, against 2 the four weeks before. 0.0% of all new papers.

Direction-Flipped Influence Audits Reveal Hidden Structure in Moral Choices of LLMs
Feb 26, 2026Phil Blandfort, Tushar Karayil, Alex McKenzie +3Language Model Bias Evaluation Moral Reasoning in Language Models
Predictably Weird · Independent · AE Studio

Feb 26, 2026J/K move · Enter open · S save

Direction-Flipped Influence Audits Reveal Hidden Structure in Moral Choices of LLMs

Phil Blandfort, Tushar Karayil, Alex McKenzie, Urja Pawar+2

Predictably Weird · Independent · AE Studio

Language Model Bias Evaluation Moral Reasoning in Language Models Evaluation Awareness in Language Models

Open paper AskPDF arXiv

Moral benchmarks for LLMs typically score models on context-free prompts, implicitly treating the measured choice rate as stable. We test this assumption with a direction-flipped influence audit: for each scenario, we compare a baseline prompt with matched cues steering toward option A or option B. Across a trolley-problem-style moral triage task, BBQ, and DailyDilemmas, and across five LLM families with and without reasoning, short contextual cues shift per-condition choice rates by 12-18 percentage points on average. These shifts reveal structure that baseline scores miss: roughly 40% of baseline-neutral triage and BBQ conditions exhibit directional asymmetry under influence, and a meaningful share of significant effects backfire, moving opposite the cue's intended direction. In follow-up probes, models often recognize the cue while denying that it affected their choice. Among significant backfire trials, this stated-vs.-revealed inconsistency appears in 78% of cases. Reasoning does not eliminate contextual sensitivity but reshapes it: social-pressure cues such as user preference and emotional appeal weaken across benchmarks, while few-shot demonstrations strengthen sharply on both triage and BBQ. We recommend direction-flipped influence pairs as a standard complement to context-free moral-bias evaluation, and release the harness and data to make such audits routine.