yaumeng

• yaumeng 17 Jul 2026 2:40 UTC

  1 point

  0

  in reply to: Yair Halberstadt’s comment on: LLM CoTs re­main mon­i­torable when be­ing un­faith­ful re­quires computation

  Another factor is the higher inference cost for longer CoTs. OpenAI’s Monitoring Monitorability paper describes this as a “monitorability tax”.

Effi­ciently De­tect­ing Hid­den Rea­son­ing with a Small Pre­dic­tor Model