Xingkai Yu (俞星凯) | Chlience Paper Archive

归档 Sep 10, 2026

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。

待审阅 2026-09-10-deepseek-v4-1-flash-kv-cache-compression KV Cache Sparse Attention Multimodal Model

归档 Jun 28, 2026

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。

待审阅 2026-06-27-dspark-confidence-scheduled-speculative-decoding Speculative Decoding Multi-Token Prediction Serving Runtime

归档 Jun 24, 2026

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。

待审阅 2401.06066-deepseekmoe-expert-specialization MoE Architecture

归档 Jun 24, 2026

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。

待审阅 2601.07372-conditional-memory-engram-scalable-lookup Memory Architecture MoE Architecture

归档 Feb 06, 2026

DeepSeek R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

用 outcome based RL 从强 base model 诱导 long CoT、自验证与策略切换，再通过 SFT / RL / distillation 转为可用模型。

待审阅 2501.12948-deepseek-r1-rl-reasoning Reasoning RL Reward Modeling RL Algorithm

Related Notes

DeepSeek V4.1 Flash: Pushing the Limits of KV Cache Compression

DSpark: Confidence Scheduled Speculative Decoding with Semi Autoregressive Generation

DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models

Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models

DeepSeek R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning