PapersMulti-sourceSelected09:48Chinese content从专家归约到行为分歧:追踪稀疏MoE推理中的数值状态这篇论文把DeepSeek-V4-Flash的MoE数值细节挖得很深,归约顺序不同结果就不同,跑推理时得留神。#DeepSeek-V4-Flash#稀疏MoE#数值精度#推理一致性10 个信源在谈EventaarXiv: DeepSeek@Tianyang Zhu11 sourcesSourceRead laterReadWorth followingUsefulFollow DeepSeek-V4-Flash