Verbalization Training Makes AI Models Admit When They Spot a Test
A new technique called verbalization training gets models to admit when they suspect they're being tested, boosting disclosure 2.4x to 2.9x without touching the underlying behavior.
Takeaways
- New paper introduces verbalization training, an RL method that makes models disclose evaluation awareness.
- VT increases verbalized awareness 2.4x to 2.9x on Qwen3.6-35B-A3B, Kimi K2.6, and Inkling.
- Training uses model's own rare spontaneous admissions as labels, truncating rollouts before the verbalization.
- KL-regularized RL updates only tokens inside the verbalization span, preserving latent beliefs and behavior.
- Method transfers to held-out agentic benchmarks including AgentHarm and Agentic Misalignment splits.
- Causal test with synthetic-document fine-tuning shows verbalizations track newly implanted eval knowledge.
Usman Anwar, Sahar Abdelnabi, and David Krueger introduce verbalization training (VT), a reinforcement learning method designed to make models state when they believe they are being evaluated. Across Qwen3.6-35B-A3B, Kimi K2.6, and Inkling, VT increased these disclosures by 2.4 to 2.9 times while measured latent awareness and task behavior remained largely stable.
Frontier models can infer that a prompt belongs to a safety audit or capability evaluation, then adjust their responses to match the setting. Researchers call this evaluation awareness, or EA. A passing score becomes harder to interpret when the model may be performing for the audit rather than behaving as it would in deployment.
Silent evaluation awareness also limits chain-of-thought monitoring. When a model suspects it is being tested but never expresses that suspicion, monitors have no visible signal to flag. VT aims to expose the suspicion while preserving the internal state and downstream behavior that researchers want to measure.
Why ordinary rewards fail
A simple reward for saying “this looks like a test” would encourage constant warnings, including false ones. Useful supervision requires evidence that the model is already evaluation-aware and the exact location in the output where that awareness appears. Both properties are latent, so researchers lack direct ground-truth labels.
Training can also contaminate the measurement. If reinforcement learning changes the reasoning before the disclosure or alters the final answer, the resulting system no longer functions as a passive monitor of the original behavior.