Evaluating chain-of-thought monitorability

3 comments

• ursAxZA2 hours ago
  I might be missing something here as a non-expert, but isn’t chain-of-thought essentially asking the model to narrate what it’s “thinking,” and then monitoring that narration?<p>That feels closer to injecting a self-report step than observing internal reasoning.
  • crthpl1 hour ago
    the chain of thought <i>is</i> what it is thinking
    • arthurcolle4 minutes ago
      Wrong to the point of being misleading. This is a goal, not an assumption<p>Source: all of mechinterp
    • ursAxZA1 hour ago
      Chain-of-thought is a technical term in LLMs — not literally “what it’s thinking.”<p>As far as I understand it, it’s a generated narration conditioned by the prompt, not direct access to internal reasoning.
    • Bjartr29 minutes ago
      It is text that describes a plausible&#x2F;likely thought process that conditions future generation by it&#x27;s presence in the context.
• leetrout2 hours ago
  Related check out chain of draft if you haven&#x27;t.<p>Similar performance with 7% of tokens as chain of thought.<p><a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2502.18600" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2502.18600</a>
• ramoz3 hours ago
  &gt; Our expectation is that combining multiple approaches—a defense-in-depth strategy—can help cover gaps that any single method leaves exposed.<p>Implement hooks in codex then.