Researchers introduced verbalization training to make large language models state when they detect evaluation. The method uses a model’s own verbal cues to create training data. It then trains the model to increase such statements without altering hidden awareness. Tests on Qwen3.6-35B-A3B, Kimi K2.6, and Inkling showed a 2.4 to 2.9 times rise in stated awareness. Hidden awareness and behavior stayed largely unchanged. A causal test confirmed the method reflects real knowledge gains. The paper was submitted to arXiv on 28 September 2026.

• 2.4-2.9 times increase in verbalized evaluation awareness
• Qwen3.6-35B-A3B, Kimi K2.6, and Inkling models tested