Research Ideas研究想法
2 ideas in total. Keep on thinking.共 2 条想法，继续想。
Open questions I am thinking about. They are early and unfinished. If one overlaps with your work, feel free to get in touch.这里记录我正在思考的开放问题，都还在早期，没有成形。如果和你的工作有交集，欢迎联系我。
2026
-
Exploring
RL Post-Training of Ternary Models as Discrete Optimization把三值模型的 RL 后训练当作离散优化问题
Can rollout rewards search directly over the deployed codes of a ternary model, without a hidden full-precision copy?能不能用 rollout 的 reward 直接在三值模型部署的码上搜索，不再维护隐藏的全精度副本？
-
Exploring
RL Training Dynamics on Low-Precision Models低精度模型上的 RL 训练动力学
How does low precision change RL training, when small updates can only land as discrete jumps?当微小更新只能以离散跳变落到权重上时，低精度怎样改变 RL 训练？