LessWrong: looping boosts no-CoT reasoning far more than general skills in NVIDIA's Dory 4B
CFGeek · x · 2026-11-22
A LessWrong study tests the worry that looped transformers are harder to monitor, using NVIDIA's public looped transformer Dory 4B.
- Result 1: Adding loops (trained at 4, still gains up to 8 at inference) substantially boosts no-CoT reasoning (NCRI) while barely moving general capability (mini-ECI) or bits-per-byte — looped models may be harder to monitor than standard transformers of similar capability
- Result 2: Dory shows near-zero CoT controllability at every loop count, limiting conclusions about controllability
The finding is in tension with claims that looped-transformer interpretability is no harder than very deep models.
More from AGI Musings
- Palmer Luckey on weapons AI: 'If I were putting ChatGPT in the head of my missile, you'd be right to be terrified' — justin_hart · 2026-11-22
- Podcast: How a Brain-Computer Interface Will Change Humanity — dr1337 · 2026-11-22
- AI poster teortaxesTex 'makes it to the boomersphere' via Ross Douthat — teortaxesTex · 2026-11-22
- Self-replicating ambient AI actuators as techno-capitalism's archetypal end — LesaunH · 2026-11-22
- Compilation of self-proclaimed AI experts failing to foresee AI's rapid progress — Chris_Armstrong · 2026-11-22
- AI labs' job posts reveal the next breakthrough: endless data harvesting and RL environments — DuaneJRich · 2026-11-22