LEV: Looped decision model lifts 2.6B accuracy 63.2% to 76.2% with adaptive compute
jiqizhixin · x · 2026-11-19
Researchers from Lehigh, UBC and the Vector Institute introduce LEV, a looped model for selecting among explicit options that spends more compute on harder questions.
How it works
- Each pass reuses the same pretrained layers, with a shared pointer head re-scoring options at every depth; a label-free stopper predicts when the current choice will match the deepest pass and exits early.
- Inference adapts depth per request, with some requests exiting at step 1.
Results
- On 231 public JevBench items, the 2.6B model goes from 63.2% accuracy at one pass to 76.2% at four passes, using the same checkpoint.
- Adaptive stopping cuts loop compute by 15–24% with under one point of accuracy loss versus fixed-depth inference; serving latency would need separate measurement.
More from Research
- Aurora2-400M-A20M: a sparse MoE pretrained from scratch in 36 hours on one RTX Pro 6000 — Tall_Abrocoma_3533 · 2026-11-19
- Researcher calls METR/Anthropic '2x speedup' RSI analysis statistically dubious — jessi_cata · 2026-11-19
- Linear Probes Could Reveal Which Persona a Base Model Is Playing, Researchers Suggest — cephaloform · 2026-11-19
- ETH's Anymal robot completes 10km autonomous alpine hike with VLM-guided ROAM system — ChongZzZhang · 2026-11-19
- OpenVax codebase grows from 78K to 565K lines in 15 months — iskander · 2026-11-19
- François Fleuret: decoder transformer's elegance — no trade-off hyperparameters — francoisfleuret · 2026-11-19