Is Evaluation Awareness Just Format Sensitivity? Limitations of Probe-Based Evidence under Controlled Prompt Structure
This reveals limitations in probe-based evidence for evaluation awareness, which is an incremental concern for researchers in interpretability and model evaluation.
The study tested whether linear probes used to detect evaluation awareness in large language models actually reflect benchmark format rather than context, finding that probes primarily track structural artifacts and fail to generalize to free-form prompts.
Prior work uses linear probes on benchmark prompts as evidence of evaluation awareness in large language models. Because evaluation context is typically entangled with benchmark format and genre, it is unclear whether probe-based signals reflect context or surface structure. We test whether these signals persist under partial control of prompt format using a controlled 2x2 dataset and diagnostic rewrites. We find that probes primarily track benchmark-canonical structure and fail to generalize to free-form prompts independent of linguistic style. Thus, standard probe-based methodologies do not reliably disentangle evaluation context from structural artifacts, limiting the evidential strength of existing results.