New FBT study: latent chains stay unused until RL length pressure forces them on
JohnCLangford · x · 2026-11-08
- A new blog post investigates how effectively the Full Bandwidth Transformer (FBT) uses its latent chain of reasoning.
- Key finding: FBT still prefers explicit CoT, but uses latents more when pressured with a length penalty.
- John Langford: the post deeply addresses RL on an FBT; some problems show strong latent use enabling extrapolation, while for many problems FBT avoids latents entirely — new methods to encourage latent use may unlock more extrapolation.
- The post also shares tricks for doing on-policy RL with FBT.
More from Research
- Mathematician publishes complete math roadmap for ML: linear algebra, calculus, probability — TivadarDanka · 2026-11-08
- Nous Research's Propaganda_Eval Dataset Trends on Hugging Face — NousResearch · 2026-11-08
- Grafted artificial neurons lift a 135M model's multiplication accuracy from 21% to 83% — CatAstro_Piyush · 2026-11-08
- Annotating 2,785 traces: researcher jokes about the grind of modern AI research — santiviquez · 2026-11-08
- Solar compiler compiles in 4.8s vs solc's 2h, cuts fuzzing RAM 4.7x in Echidna test — banteg · 2026-11-08
- Cornell professor recruiting 1-2 PhD students on SE×AI quality assurance for Fall 2027 — AI4Code · 2026-11-08