这个意外发现很有意思：混一点点推理数据，不改变推理过程，象棋AI就变强了。适合研究模型训练技巧的人。
开发者Amjad Masad发现，训练chess LLM从棋盘局面到走子对时遭遇收益递减。意外中，一个“先思考再走子”分支虽然产生幻走，但混合约8%的推理轨迹到纯走子数据后，在相同预算下效果超过纯走子数据。模型内化了推理过程，但推理时不显式使用。新部署的象棋引擎估计约1200 Elo，目标是达到2000+，且保持小规模微调LLM、无引擎辅助的约束。
hit massive diminishing returns training my chess LLM on board→move pairs. then an accident: a "th...
hit massive diminishing returns training my chess LLM on board→move pairs. then an accident: a "think before you move" branch failed (hallucinated moves) but mixing just ~8% of those reasoning traces into plain board→move training beat pure move data at equal budget. the model internalizes the thinking without ever doing it at inference. Amjad Masad @amasad Just deployed a new chess engine; we're closing in on an estimated 1200 Elo. The goal is to get it to 2000+, while keeping the constraints: 1. One small finetuned LLM (no custom pretraining or architecture) 2. The model has to produce the moves with no chess engine assistance If you relax these constraints, it gets much easier. 🔗 View Quoted Tweet 💬 5 🔄 3 ❤️ 20 👀 3255 📊 6 ⚡