tracked author
Wenfeng Liang (梁文锋)
DeepSeek founder / senior organization signal and repeated submitter-author across DeepSeek-V2, DeepSeek-V3, DeepSeek-R1, DeepSeek-V3.2, DeepSeek-V4 and Engram. Public social-profile evidence is not strong enough for a personal X account.
Related Notes
按论文归档时间排序，展示该作者在本站已经出现的材料。
DeepSeek V4.1 Flash: Pushing the Limits of KV Cache Compression
DeepSeek V4.1 Flash 将因果编码器与解码器分工、跨层稀疏 KV 复用、四位全局缓存和近似窗口重放组合起来，把长输入的预填充主干计算近似减半、全局 KV 降至每 token 890 字节，并在发布方评测中显著提升多项 agent 能力；端到端服务收益和近似重放的极端条件可靠性仍缺少充分公开验证。
DSpark: Confidence Scheduled Speculative Decoding with Semi Autoregressive Generation
用 Markov head、置信度校准和硬件感知前缀调度，把并行 drafter 推进生产 serving。
DeepSeek V3.2: Pushing the Frontier of Open Large Language Models
用 DSA 降低 128K attention 成本，再以四层 GRPO consistency control 和 85,267 条真实/合成 agent tasks 扩展 reasoning 与 tool use。
DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models
用细粒度专家切分和共享专家隔离提高 routed expert 的专门化程度。
Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models
用 hashed N gram lookup 和 context aware gating 增加可离线扩展的 conditional memory。
DeepSeek V3 Technical Report
把 auxiliary loss free MoE balancing、MTP、FP8 与 DualPipe 组合成高效大规模训练系统。
DeepSeek V2: A Strong, Economical, and Efficient Mixture of Experts Language Model
把 per head K/V cache 改成由 hidden state 下投影得到的共享 KV latent，并用 projection absorption 与 decoupled RoPE 避免恢复历史 K/V。