songdj/Lev-1.4B
Updated • 21
Lev: Thinking Fast and Deep with Looped Decision Models. One checkpoint, multiple depths: 3 models + LevBench, training data and results.
Note Ouro-1.4B backbone, depths 1-4. JevBench 58.0% -> 68.0%.
Note arXiv:2610.20951
Note Ouro-2.6B backbone, depths 1-4. JevBench 63.2% -> 76.2%.
Note Nanbeige4.2-3B backbone, depths 1-2. JevBench 69.3% -> 72.7%.
Note Evaluation: 5 tiers, development + locked test (rights-filtered subset).
Note Training: S31 / S60 / S120 mixtures and recipes (rights-filtered subset).
Note Per-depth outputs, calibration and latency; no question text.