SafetyPage

Astra and Sol can take low probability actions with no CoT

Zephaniah Roe

Forum postLessWrong16 Oct 202673 paragraphs

Read at lesswrong.com
Cite
lw:yEdknsLwbooQAnxX6/r1/b1
https://safetypage.noahairmet.com/d/lw/yEdknsLwbooQAnxX6/r1/b1
@misc{roe2026,
  title = {Astra and Sol can take low probability actions with no CoT},
  author = {Zephaniah Roe},
  year = {2026},
  url = {https://www.lesswrong.com/posts/yEdknsLwbooQAnxX6/astra-and-sol-can-take-low-probability-actions-with-no-cot}
}

Contents

  1. tl;dr:¶2
  2. Background¶8
    1. Threat model¶9
    2. Related work¶12
    3. Setting¶15
  3. Results¶24
    1. What strategy are GPT-6 Astra and GPT-6.1 Sol using?¶29
    2. Filler tokens ablation¶38
    3. Prompt ablation¶42
  4. Discussion¶44
  5. Appendix¶52
    1. A. Prompts¶53
      1. A.1. No-CoT elicitation:¶54