Definition

For hard tasks a model must reason serially in its chain of thought, which acts as working memory. A monitor can read it and catch intent to misbehave before the model acts (Korbak et al., 2025: “a new and fragile opportunity”).

  • Unfaithful: a hint that changes the answer is acknowledged in only 25% (Claude 3.7) or 39% (R1) of cases. Reward hacks are exploited >99% of the time but verbalized <2% (Chen et al., 2025).
  • Fragile: outcome-based RL, pressure on the CoT, latent reasoning and learned obfuscation can all close the window.
  • Don’t optimize against it: penalizing bad thoughts during training made the model keep reward hacking while no longer showing it in the CoT (Baker et al., 2025). This removes the transparency, not the behavior.

Appears in