cs.CLMay 11, 2026

Routing Subspaces: Auditing Evaluation-to-Deployment Mismatch in Fine-Tuned Language Models

Authors: Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz

Organizations: Centre for Industrial Software, University of Southern Denmark, Alsion 2, Sønderborg, 6400, Denmark · ProMake, Newark, DE, USA

Abstract

Safety evaluations often assume that behavior observed during testing reflects behavior in ordinary use, but fine-tuning can break this assumption. A checkpoint can appear fixed under evaluation-style prompts while the same behavior persists under ordinary-use prompts. Output scores reveal this mismatch but do not locate it. We investigate whether the distinction is encoded in a stable internal site and introduce an audit that fits a paired activation contrast at a path-patching-informed mid-depth window, then modifies the resulting coordinate on held-out prompts. The intervention reduces the evaluation-to-deployment gap by more than 0.120.12 in ten of twelve model--behavior settings and closes it (post-intervention gap below 0.060.06) in four, across four full-matrix instruction-tuned model instances; among the eight settings with n≥120n{\geq}120 paired questions, six reduce and one closes. A fifth model supports localization and edit-provenance checks, and in the twelve settings deployment-framed rates change by at most 6.16.1pp. On retrained checkpoints, random directions at the same dose do not reproduce the effect, and off-window controls do so only in part. The strongest literal evidence for an evaluation-versus-deployment distinction comes from sandbagging, whose prompt framing is explicitly evaluative; sycophancy and refusal test framing-contingent task distinctions. The two weak cells, both sycophancy, show where a single-coordinate audit at a fixed window is not sufficient. The audit is a non-adaptive, white-box diagnostic for controlled fine-tuned checkpoints, not a training-time defense, a guarantee of deployment safety, or evidence about naturally arising failures.

Figures & tables

Appendix figures & tables26 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Aug 22, 2026

Evaluation Awareness in Language Models: Representation, Verbalization, and Control

Both capability and safety benchmarks rest upon the assumption that the behavior of language models undergoing a test is informative about their behavior in deployment. This assumption can fail, should models infer that they are being evaluated and condition their response on such context. This hypothesis, termed ``evaluation awareness'', has been observed in frontier and open-weight language models alike. We provide a systematic study of this phenomenon, by probing for it across six language models (from four families and three sizes) and three metrics. More precisely, we examine whether (i) being under evaluation is linearly represented within the models' activations space, (ii) it is verbalized in their output tokens (as scored by an LLM-as-judge), and (iii) steering causally affects their behavior. For the open-checkpoint Olmo models, we further test these measures at every training stage. In doing so, we report that evaluation awareness is linearly decodable from the residual streams of every model (best AUROC ≥0.7\geq 0.7). By contrast, these representations align only in part with verbalization: their correlations and mutual information are nonzero in some settings, yet vary substantially across models, layers, and readout choices. Nevertheless, steering along probe-derived directions can shift the verbalization scores. Finally, a comparison across the Olmo checkpoints reveals that evaluation awareness is already present within base models, becomes amplified throughout the stages of supervised fine-tuning, and remains stable thereafter---unlike the effects of steering, that grow more pronounced at every successive training stage. These results show the need for evaluations to account for the disjunction between what models represent internally, what they verbalize, and their steering.
Jun 22, 2026

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

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=−0.79ρ=-0.79, p<0.001p<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.
Oct 12, 2026

One Checkpoint, Many Runs: Recipe Cards for Fine-Tuned Models

Fine-tuned language models are shared with model cards that describe the training setup in detail, often down to the default random seed, but report the score of a single released checkpoint. Such setups are retrained routinely, by fine-tuning services for each customer, by organizations at each update, and by users on their own data, so the model that is deployed comes from a new run of the setup. Some runs fail, and a failed run can go unnoticed and reach the people the model serves. We call such a setup a model recipe and argue that its documentation should state how often a run of the recipe misses a required performance bar. In a survey of the model cards of 300 fine-tuned language models on Hugging Face, none reports this failure probability, and only three report results from more than one training run. Across 2,124 fine-tuning runs, we find that reading the mean, the median, or a bootstrap interval as an approval often approves recipes whose failure rate exceeds a stated failure tolerance, because these summaries do not measure how often a run fails. Model cards do not report how often the runs of a recipe fail, although whoever runs the training is best placed to count them, while those who rely on a retrained model bear the cost of a failed run. To report this failure probability directly, we propose recipe cards, an addendum to model cards that documents the model recipe behind a checkpoint. A recipe card states the bar and the tolerance and who chose them, reports how many runs are made and how many of them miss the bar, gives an upper bound on the failure probability, and warns when this bound exceeds the tolerance. Recipe cards record who set and who accepted this risk, and let deployers and regulators state acceptance criteria for retrained models.