Definition

Chain-of-thought is intermediate reasoning text that models generate (or are prompted to generate) as a scratch pad while solving tasks. It is privileged as “the model’s own thinking,” which makes style-spoofed forgeries dangerous (cot-forgery).

Key Points

  • Used for multi-step reasoning and agent planning
  • Security implication: if model identifies CoT by style not tags → role-confusion
  • ICML 2026 work shows forged CoT can raise jailbreak success to ~60% (2026-07-30-cot-forgery-arxiv)

Sources