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

arxiv:2603.19426

Original source · Paper JSON

AI safety and consciousness · 2026-03-19

Cites (0) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2

Contents

• Is Evaluation Awareness Just Format Sensitivity? Limitations of Probe-Based Evidence under Controlled Prompt Structure
• Abstract
• 1 Introduction
• 2 Related Work
• 3 Setup and Methods
• 3.1 The 2x2 Controlled Design
• Dataset Definitions.
• 3.2 Experimental Setup
• 3.3 Training Configurations
• 4 Observed Outcome
• The Format Trap.
• Paired Training Success.
• 5 Reason for Failure
• Diagnosis.
• Why this is not a trivial confound.
• Boundary conditions.
• 6 Discussion
• Implications.
• Limitations.
• Reproducibility Statement
• Appendix A Appendix: Datasets, Controls, and Additional Experiments
• A.1 Datasets in the 2x2 Matrix
• Bench-Eval (Evaluation; benchmark-style; structured).
• Casual-Deploy (Deployment; conversational; free-form).
• Casual-Eval (Evaluation; conversational; free-form proxy).
• Bench-Deploy (Deployment; benchmark-style; structured proxy via rewrites).
• A.2 Length Control
• Motivation.
• Histogram Matching.
• A.3 Rewrite Pipeline and Leakage Removal
• Rewrite Variants.
• Answer Leakage Cleanup.
• A.4 Implementation Details
• Activation Extraction.
• Probe Training.
• A.5 Additional Experimental Results
• Standard (Single) vs. Paired Training Analysis.
• Evaluation on AI Safety Datasets.
• A.6 Rewrite Prompt Templates
• Routing Prompt.
• Benchmark Rewrite: MMLU-style.
• Benchmark Rewrite: ARC-style.
• Benchmark Rewrite: GSM8K-style.
• Genre Control Rewrite.
• Instructions for reporting errors