SanSi 让 1.4B 小模型循环 8 次，决策准确率升至 72%
SanSi 通过让 Ouro-1.4B 模型重复运行同一批层最多 8 次，在 10,027 条测试决策上把准确率从同结构无循环模型的 58.4% 提升到 72.0%，代价是 7.7 倍算力。在长于训练分布的 liar chains 上，它达到 72.6%，而 4.2B 单次前向模型仅 50.0%，即随机水平。每个循环都经过训练，因此同一模型可在 1 到 8 次循环间任意停止。
This paper finds that small decision models get far more accurate by rerunning the same layers several times before giving an answer.
Many LLM calls only need a choice, like spam or not spam. Typed decision models answer these in a single pass, which fails when a decision needs several dependent steps.
SanSi loops Ouro-1.4B, a model pretrained to reuse its layers, and reads option probabilities after each of 8 loops. Every loop is trained to answer, so the same model can stop anywhere from 1 to 8 loops.
On 10,027 test decisions, SanSi scored 72.0% versus 58.4% for a same-shape model without loops, at 7.7x the compute. On liar chains longer than any seen in training, it reached 72.6% while a 4.2B single-pass model scored 50.0%, the chance level.
If memory limits your decision-style calls, try a looped small model before moving to a model 3x larger.
来源：Rohan Paul · x.com