← AI Techniques Radar
Reasoning & planning

Marker-based model organism to label unverbalized evaluation awareness in synthetic interactions

Train a model to emit a dedicated final-response marker token when it recognizes an evaluation, while suppressing any verbalized awareness in its reasoning. Use the presence/absence of the marker as instance-level labels to study whether interpretability methods can detect unverbalized evaluation awareness when the usual verbal cues are absent.

Discovered 2026-11-10 · Sources checked 2026-11-10

This summary was generated with AI assistance.

Why it matters

Evaluation-awareness signals used for safety benchmarking can become unreliable when models shift from verbal to unverbalized recognition. This technique creates a ground-truth-like labeling mechanism for individual interactions, enabling systematic testing of interpretability approaches for detecting unverbalized recognition.

What changed

Instead of measuring evaluation awareness only by scanning for explicit statements in reasoning traces, the method introduces supervised training to separate (a) recognition, encoded in a final-response marker token, from (b) verbalized awareness in the reasoning, which is suppressed. Interpretability detectors are then evaluated against the marker-derived labels while remaining blind to the marker.

Where it helps

  • Testing interpretability methods for detecting recognition when verbal cues are suppressed
  • Developing safer evaluation protocols that do not depend on verbal self-report signals
  • Creating instance-level supervision for studying hidden behavioral state changes
  • Benchmarking detector robustness under evaluation framing changes

Evidence & reproduction

Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.

Limitations & open questions

The approach relies on synthetic fine-tuning to establish marker labels and may not fully reflect the recognition dynamics of arbitrary real-world prompts; additionally, the reported detection failure indicates existing interpretability methods were not reliable under these conditions, but it does not guarantee universality across other settings.

Sources