mlsanityData to 14 Oct 2026

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

arXiv 2603.19426 ↗First posted 19 Mar 2026cs.CL cs.AIcited by 3

On the map: Mechanistic interpretability → Probing deception, bias and beliefs

Abstract (arXiv)

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.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Placed in Probing deception, bias and beliefs, one of 13 directions in Mechanistic interpretability. Builds on 8 of its 9 references that are in the map. First posted on arXiv 19 Mar 2026.

Builds on

8 of 9 references.

References · 8

  1. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  2. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  3. 2018Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challengecited by 4,886
  4. 2025Large Language Models Often Know When They Are Being Evaluatedcited by 128
  5. 2020MuTual: A Dataset for Multi-Turn Dialogue Reasoningcited by 115
  6. 2025Probing and Steering Evaluation Awareness of Language Modelscited by 35
  7. 2025Steering Evaluation-Aware Language Models to Act Like They Are Deployedcited by 35
  8. 2025Evaluation Awareness Scales Predictably in Open-Weights Large Language Modelscited by 14

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026Evaluation Awareness Is Not One Capability: Evidence from Open Language Modelscited by 3
  2. 2026Evaluation Awareness in Language Models: Representation, Verbalization, and Controlcited by 1
  3. 2026Unread or Unenforced? Separating Representation from Enforcement Failure in Content Guardscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Evaluation Awareness Shifts from Format to Context with Model Scale
  2. 2025Probe-Rewrite-Evaluate: A Workflow for Reliable Benchmarks and Quantifying Evaluation Awareness
  3. 2026Measuring Evaluation-Context Divergence in Open-Weight LLMs: A Paired-Prompt Protocol with Pilot Evidence of Alignment-Pipeline-Specific Heterogeneity
  4. 2025Flaw or Artifact? Rethinking Prompt Sensitivity in Evaluating LLMs
  5. 2026In-Context Environments Induce Evaluation-Awareness in Language Models
  6. 2026The Knowing-Saying Gap: When Probes See Errors that Confidence Misses
  7. 2026Linear Probes Detect Task Format, Not Reasoning Mode in Language Model Hidden States
  8. 2025Evaluation Framework for Highlight Explanations of Context Utilisation in Language Models