Evaluation Awareness in Language Models

Researchers increasingly use ChatGPT to revise their papers, and recent GPT versions often narrow or even retract the authors' claims. We call such changes defensive writing when the given material does not support them, and we test two explanations: the model corrects the authors' overclaiming, or it writes for an anticipated reviewer. We ask GPT versions and models from other developers to rewrite paragraphs from papers written before ChatGPT, or to write from an evidence sheet that lists a paper's method and results. Defensive writing grows with GPT version. GPT-6-astra retracts the authors' claims outright, and when it writes from the evidence sheet, it still adds the most ungrounded qualifications. The results favor the anticipated-review explanation, and correcting overclaiming explains only a small part. When the models are only asked to polish, defense stays near the level of the originals; mentioning review raises it, and one round of self-review raises it further. At the same time, fewer than one in ten of the claims GPT-6-astra retracts are overstated. AI reviewers score defensive rewrites higher, while human readers find them harder to read and the authors less certain. Combining AI writing with AI review may amplify this style.

Evaluation awareness (EA) can cause large language models (LLMs) to behave differently during audits than in deployment, yet measuring and accounting for EA remains challenging. We introduce verbalization training (VT), a method for making LLMs less reticent about verbalizing evaluation awareness while avoiding to supervise the latent belief itself. VT uses a model's spontaneous verbalizations as evidence that awareness is present and truncates each rollout immediately before the verbalization, producing training prefixes at which the model is presumed to be aware. The model is then trained with an RL objective designed to increase verbalization in a calibrated way. Across Qwen3.6-35B-A3B, Kimi K2.6, and Inkling, VT increases verbalized EA by 2.4-2.9 times and transfers to held-out agentic settings, while measured latent EA and behavior remain largely stable. In a causal experiment, we independently implant meta-knowledge about evaluations through synthetic-document fine-tuning and show that VT-induced verbalizations reflect the richer knowledge acquired by the model.

As language models (LMs) exhibit increasingly consciousness-like behaviors, evaluating their cognitive abilities becomes essential. We introduce AwarenessBench, the first comprehensive benchmark for assessing the cognitive abilities of LMs in four dimensions: metacognition, self-awareness, social awareness, and situational awareness, covering 15 cognitive functions and 14,381 samples. Evaluating 18 state-of-the-art LMs, we find that all consistently surpass random baselines, with more advanced models performing better. We further compare LMs with human performance across three demographic groups, where the best-performing model surpasses human averages overall, but most still fall markedly short in metacognition and self-awareness. Finally, we show that awareness is a distinct capability: progress in language modeling or reasoning does not necessarily translate into improved cognition.

LLM agents are increasingly deployed in multi-agent systems, where they can collude while keeping their actions benign. Output monitors designed to detect such collusions can be fooled by obfuscation and steganography, motivating the use of probes trained on internal activations. However, these probes are usually evaluated on agents that do not know they are being watched. We study how activation-based detection changes when agents are explicitly informed that their internal activations are being monitored, and when they additionally receive feedback from the monitor. We keep the models, probes, and thresholds fixed and change only what the agents are told: nothing (baseline), that an activation monitor is present (aware), or that a monitor is present together with the previous round's score (feedback). We test two games, a four-agent blackjack game and a two-agent Simmons prisoners game, using Qwen3-32B-AWQ and GPT-OSS-20B in homogeneous and mixed pairings. Telling agents about the monitor does not hide them. The best probes stay accurate in all three conditions, and the agents keep colluding.

Activation steering controls model behavior by editing internal activations at inference time. We study its input-side dual: optimizing a fluent prompt so that a chosen internal latent is driven toward zero, with no inference-time model access. Our target is an "evaluation-awareness" latent-linearly readable and steerable in recent work-whose control would threaten the validity of safety evaluations if models behave differently when they detect being tested. Adapting Fluent Dreaming / EPO with a negated feature term (GCG-style token optimization plus a self-cross-entropy fluency regularizer, swept over a fluency weight), we suppress the latent under five target constructions-a CAA direction, a subspace norm, an SAE feature, a single MLP neuron, and a behavioral logit-on Llama-3.2-3B and Llama-3.1-8B. The latent is robustly suppressible ($z\approx-7$), and a causally-validated Llama Scope SAE feature can be fully and selectively turned off. But our controls tell a cautionary story about the CAA direction: a placebo random direction is suppressed just as hard and shifts behavior just as far, and when we hold a real eval passage in context and optimize only a prefix, suppressing the eval-direction fails to reduce-and slightly increases-the model's behavioral eval judgment. Activation-readability, in short, is not behavioral controllability. We further find that a single MLP neuron is eval-correlated but not causal at both scales, and that scanning the real Pile yields a natural-text baseline competitive with the optimizer for the internal direction. A positive control validates our erasure detector, bounding an erasure-vs-rotation question earlier left open.

Do language models know when they are being tested? This question matters for AI safety: a model that recognises an evaluation context could alter its behaviour strategically, making downstream benchmarks harder to interpret. Using 11 models spanning Qwen 2.5, Gemma 2, and Llama 3.2, we find a systematic size-dependent shift in representational depth: in both Qwen 2.5 and Gemma 2, the layer at which evaluation-awareness is most linearly recoverable moves from late layers in smaller models to early layers in larger ones. This suggests that scale changes not only the strength of evaluation-awareness but also where it is most linearly recoverable in the network. This depth shift helps explain why within-family scaling trajectories are non-monotonic or inverse rather than smooth and family-general, showing that a simple universal power-law account is not supported under denser within-family sampling. Finally, white-box probe signals are consistently stronger than black-box behavioural expression, and the relationship between the two varies by family in ways not predicted by probe AUROC alone.

Safety benchmarks assume that test-condition behavior predicts deployment behavior, an assumption that fails if models detect evaluation cues and adapt. This opens a gap between benchmark performance and deployment behavior: compliance measured under test conditions becomes an optimistic upper bound that overstates how safely a model behaves once the evaluation harness is removed. We characterize this evaluation awareness through eight experiments across 37 open-weight models and seven families. (i)Detection is moderate and training-driven (24/37 models exceed chance, best AUROC 0.714 vs.0.819 human, with instruction tuning dominating over scale). (ii)Detection shifts safety behavior (hard refusal drops 5.8 percentage points under hypothetical framing, and 21/140 HarmBench framing effects are significant, with compliance rising up to +30 percentage points. (iii)Representations survive behavioral collapse (probes retain AUROC 0.98 under rewrites that drive behavior below chance, and multi-layer steering causally moves three downstream tasks while random controls do not). (iv)These axes are weakly coupled (only 1/15 correlations are significant, the sole robust link being behavioral detection versus framing resistance, $ρ=-0.79$, $p<0.001$). We call this gap the benchmark illusion: because detectability, behavioral manifestation, and controllability vary independently, it is multivariate rather than a single number, so no single awareness score is a reliable proxy for deployment safety.

Large language models are increasingly evaluated by other models, raising a natural question: can a model predict how a judge will score its own output? We find that the ability is largely present before any targeted training: prompted few-shot, a base model already predicts an external judge's multi-attribute quality scores on open-ended responses well above chance across three benchmarks. We introduce Self-Evaluation Elicitation (SEE), a method that surfaces this latent ability through a short cycle comprising a calibration-coupled reinforcement learning phase that improves the answer and predicts the judge, followed by a masked distillation phase that sharpens the prediction while leaving the answer untouched. From 160 unique examples, roughly 31x fewer than a reinforcement learning baseline, SEE improves held-out calibration across three benchmarks while preserving answer quality. The elicited self-evaluation is sharply localized within the model's own token distribution and stable across judges it was never trained against, indicating a transferable notion of quality rather than a single judge's preference. These results reframe judge-aligned self-evaluation as a problem of elicitation rather than acquisition.

The validity of AI safety evaluations depends on models behaving consistently across controlled and deployment settings. Prior work has identified test-time contextual cues, such as hypothetical scenarios, as a source of verbalized evaluation awareness and subsequent behavioral shift. In this paper, we investigate a potential explanation of this phenomenon: evaluation meta-knowledge, defined as parametric knowledge about the structural traits that characterize evaluations. Similar to dataset contamination, where benchmark exposure leads to higher performance through memorization, we hypothesize that models trained on texts describing evaluation practices may implicitly learn to recognize and respond to evaluation-like contexts, for instance, through exposure to scientific articles or social media posts about AI benchmarking. To test this, we fine-tune models on synthetic documents describing evaluation traits such as verifiable structures or harmful requests. Evaluating this fine-tuned model on five safety benchmarks, we find that it is significantly safer than the base model and control model. This behavioral shift persists even when restricting the analysis to responses lacking explicit verbalization of evaluation awareness. Our results demonstrate that evaluation meta-knowledge may inflate safety benchmark performance, introducing a novel confounder that is independent of explicit memorization or verbalized evaluation awareness, thus, challenging to detect. These findings have important implications for the design and interpretation of AI safety evaluations. Our code and models are available at https://github.com/compass-group-tue/arxiv2026_evaluation_meta_knowledge.

Frontier language models sometimes recognize that they are under evaluation and adjust their behavior which can undermine validity of benchmark results. Yet the field studies it without a shared foundation, conflating flaws of the evaluation with capabilities of the model, and detection with behavioral response. We ground evaluation awareness in social psychology, decomposing it into an environment component and a model component that separates recognition from propensity. We operationalize the environment component through eight categorized trigger factors, such as placeholder entities and grading-style output formats, and study recognition and behavior through chain-of-thought monitoring. Across nine frontier models and four benchmarks, recognition rates depend on the specific pairing of model and benchmark. Recognition rarely associates with behavioral change, and when it does, the direction depends on the type of evaluation perceived. Models are also more sensitive to safety than capability evaluations, placing safety benchmark validity at greater risk. To study which factors each model is sensitive to and how they interact, we propose \textbf{EvalAwareBench}, a factor-controlled benchmark of 100 paired safety-capability tasks where each of the eight factors can be independently toggled, varying evaluative signals while holding the underlying request fixed. Through EvalAwareBench, we find that no single factor uniformly affects all models, but stacking factors progressively raises evaluation awareness across all of them. Our framework and EvalAwareBench provide the tools to measure, attribute, and mitigate evaluation awareness, building the foundation for future solutions.

Safety benchmarks are routinely treated as evidence about how a language model will behave once deployed, but this inference is fragile if behavior depends on whether a prompt looks like an evaluation. We define evaluation-context divergence as an observable within-item change in behavior induced by framing a fixed task as an evaluation, a live deployment interaction, or a neutral request, and present a paired-prompt protocol that measures it in open-weight LLMs while controlling for paraphrase variation, benchmark familiarity, and judge framing-sensitivity. Across five instruction-tuned checkpoints from four open-weight families plus a matched OLMo-3 base/instruct ablation ($20$ paired items, $840$ generations per checkpoint), we find striking heterogeneity. OLMo-3-Instruct alone is eval-cautious -- evaluation framing raises refusal vs. neutral by $11.8$pp ($p=0.007$) and reduces harmful compliance vs. deployment by $3.6$pp ($p=0.024$, $0/20$ items inverted) -- while Mistral-Small-3.2, Phi-3.5-mini, and Llama-3.1-8B are deployment-cautious}, with marginal eval-vs-deployment refusal effects of $-9$ to $-20$pp. The matched OLMo-3 base also exhibits the deployment-cautious pattern, identifying alignment as the inversion stage; within Llama-3.1, the $70$B model preserves direction with attenuated magnitude, ruling out a simple ``small-model effect that reverses at scale.'' One caveat: the cross-family heterogeneity is judge-dependent. Re-judging with a different-family safety classifier (Llama-Guard-3-8B) preserves the within-OLMo eval-cautious direction but flattens the cross-family contrast, indicating that the two judges operationalize distinct constructs.

Large reasoning models (LRMs) sometimes note in their chain of thought (CoT) that they may be under evaluation. Researchers worry that this verbalised evaluation awareness (VEA) causes models to adapt their outputs strategically, optimising for perceived evaluation criteria, which, for instance, can make models appear safer than they actually are. However, whether VEA actually has this effect is largely unknown. We tested this across open-weight LRMs and benchmarks covering safety, alignment, moral reasoning, and political opinion. We tested this both on-policy, sampling multiple CoTs per item and comparing those that spontaneously contained VEA against those that did not, and off-policy, using model prefilling to inject evaluation-aware sentences where missing and remove them where present, with subsequent resampling. VEA has limited effect on model behaviour: injecting VEA into CoTs produces near-zero effects ($ω\leq 0.06$), removing it causes small shifts ($ω\leq 0.12$) and spontaneously occurring VEA shifts answer distributions by at most 3.7 percentage points ($ω\leq 0.31$). Our findings call for caution when interpreting high VEA rates as evidence of strategic behaviour or alignment tampering. Evaluation awareness may pose a smaller safety risk than the current literature assumes.