推荐exhaze · 2026-07
“felt you might find this paper on reasoning models trained on physical world video data becoming better at general reasoning interesting:”
推荐它,讲在物理世界视频上训练可能提升推理模型的通用推理
让语言模型借物理引擎模拟结果来做物理推理
仅用文本训练的语言模型缺乏对现实世界的接地体验,因此在物理推理上会犯明显错误。Mind's Eye 的做法是:针对物理推理问题,用计算物理引擎(DeepMind 的 MuJoCo)模拟可能的结果,再把模拟结果放进输入,让语言模型据此推理。在包含 39 个任务的物理对齐基准上,平均零样本绝对准确率提升 27.9%,少样本提升 46.0%,配上它的小模型能达到大 100 倍模型的水平。
评论者把它当作“把大模型接上物理模型”的例子,说明语言模型可借 MuJoCo 之类的模拟器完成接地推理。
“felt you might find this paper on reasoning models trained on physical world video data becoming better at general reasoning interesting:”
“An example that's even more impressive than "wrap Wolfram Alpha" has already been demonstrated: LLMs plugged into Physics models (MuJoCo) at Google”