LEV: Thinking Fast and Deep with Looped Decision Models
LEV: Thinking Fast and Deep with Looped Decision Models
Lev 是一款 System 1.5 決策模型,透過重複執行循環骨幹網絡並在每輪後為選項評分,無須解碼思維鏈即可在不同深度作答。於 JevBench 的 231 道公開題目中,Lev-2.6B 從執行一輪時的 63.2% 提升至四輪時的 76.2%,增加 13.0 個百分點。
Can a decision model think longer without writing a longer answer?
Lev is a System 1.5 decision model: it reruns a looped backbone and scores the options after every loop, without decoding a chain of thought, so one checkpoint answers at any depth. On JevBench's 231 public items, the same Lev-2.6B goes from 63.2% at one loop to 76.2% at four (+13.0 pp), mostly on standard and hard questions.
A small stopper fitted on unlabelled requests saves 15-24% of loop computation for 0.3-0.8 pp (offline analysis). On our locked test, this adaptive gain holds for 1.4B and 3B, not for 2.6B.
📄 Paper: https://arxiv.org/abs/2610.20951
🌐 Project: https://openlair.github.io/lev/
💻 Code: https://github.com/OpenLAIR/lev
🤗 Models & data: https://huggingface.co/collections/songdj/lev
The video replays recorded readouts from our runs: an easy question stops after one loop; a hard one changes its answer from 4 to the correct 6 by loop four.
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co