SafetyPage

Alignment Forum and LessWrong · Forum post · LessWrong

Astra and Sol can take low probability actions with no CoT

Zephaniah Roe · 16 Oct 2026

Read at lesswrong.com ↗Read at sourceNo licence statement found

Outline · 73 paragraphs

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