Research

Paper

TESTING March 19, 2026

Is Evaluation Awareness Just Format Sensitivity? Limitations of Probe-Based Evidence under Controlled Prompt Structure

Authors

Viliana Devbunova

Abstract

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.

Metadata

arXiv ID: 2603.19426
Provider: ARXIV
Primary Category: cs.CL
Published: 2026-03-19
Fetched: 2026-03-23 16:54

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.19426v1</id>\n    <title>Is Evaluation Awareness Just Format Sensitivity? Limitations of Probe-Based Evidence under Controlled Prompt Structure</title>\n    <updated>2026-03-19T19:39:35Z</updated>\n    <link href='https://arxiv.org/abs/2603.19426v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.19426v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>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.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <published>2026-03-19T19:39:35Z</published>\n    <arxiv:comment>10 pages, 5 tables, 2 figures. Accepted at ICLR 2026 Workshop \"I Can't Believe It's Not Better\"</arxiv:comment>\n    <arxiv:primary_category term='cs.CL'/>\n    <author>\n      <name>Viliana Devbunova</name>\n    </author>\n  </entry>"
}