Library of Layers: models route transformer block order per input to control algorithms
jonasgeiping · x · 2026-11-04
Library of Layers introduces models that adapt layer order per input, routing over whole transformer blocks from a shared layer library. This opens a new control axis: the model's execution path. The team shows that merely changing layer order switches which algorithm the model uses to solve mazes and sudoku.
Related event: Library of Layers: Reworking Transformer Layer Order at Inference Time(3 posts)→
More from Research
- Team burns 10B tokens in 2.5 days to build Omniscience, a city-scale world model — Kuprel · 2026-11-04
- Berkeley researchers use RL to improve robot data collection with Data Inoculation — berkeley_ai · 2026-11-04
- Map2Video uses maps and street view imagery to direct AI video of real locations — ryosuzk · 2026-11-04
- Xudong Sun wins honorable mention for SIGOPS Dennis M. Ritchie Dissertation Award — LingmingZhang · 2026-11-04
- First in-human gene correction fixes one-letter DNA typo in patient's liver — theimposingshadow · 2026-11-04
- 15-year-old's attention kernel survives 32k tokens where standard PyTorch OOMs, 11x faster — vankownlaw · 2026-11-04