331 posts in total. Keep on posting.

2026

• 06-11 中

  MegaScale：ByteDance 如何在 12,288 块 GPU 上实现 55% MFU 的大规模 LLM 训练

  MegaScale 是 ByteDance 用于超大规模 LLM 训练的生产系统，通过算法-系统协同设计、通信计算重叠、算子优化和深度可观测性，在 12,288 块 GPU 上实现了 55.2% 的 Model FLOPs Utilization，比 Megatron-LM 提升 1.34 倍。

• 06-10 EN

  KeepKV: Lossless KV Cache Compression via Electoral Votes and ZIP-Merging

  KeepKV introduces Electoral Votes and Zero Inference-Perturbation Merging to achieve single-step lossless KV cache compression, provably fixing the Attention Sag problem that plagues all prior merging methods.

• 06-10 中

  KeepKV：用「选举票」机制和零扰动合并实现无损 KV 缓存压缩

  KeepKV 提出了「选举票」机制和零推理扰动合并（ZIP-Merging），在数学上证明了单步无损 KV 缓存压缩，从根本上解决了所有现有合并方法都存在的「注意力衰落」问题。

• 06-09 EN

  VAPO: Value-Augmented Proximal Policy Optimization for Long-CoT Reasoning

  VAPO revives value-model-based RL for LLM reasoning by introducing Length-adaptive GAE and a suite of complementary techniques, reaching 60.4 on AIME 2024 with Qwen2.5-32B — outperforming DAPO by more than 10 points in under 5,000 training steps.

• 06-09 中

  VAPO：面向长链推理的价值增强近端策略优化

  VAPO 通过引入长度自适应 GAE 以及一套互补技术，让基于价值模型的强化学习重新超越了无价值模型方法，在 Qwen2.5-32B 上以不足 5000 步达到 AIME 2024 得分 60.4，比 DAPO 高出 10 分以上。

• 06-08 EN

  ExpWeaver: How LLM Agents Learn from Past Experience in Latent Space

  ExpWeaver replaces text-based experience retrieval with latent-space RAG — encoding past agent trajectories as dense hidden-state vectors and retrieving them at every decoding step via cross-attention, achieving SOTA on 12/13 tasks with 1.5-2x better token efficiency.

• 06-08 中

  ExpWeaver：LLM 智能体如何在隐空间中从经验中学习

  ExpWeaver 用潜空间 RAG 替代文本检索——将智能体的历史轨迹编码为稠密隐状态向量，在每个解码步骤通过交叉注意力检索并融合，在 12/13 个任务上取得 SOTA，同时将词元消耗降低 1.5-2 倍。

• 06-07 EN

  SlidingServe: SLO-Aware Sliding-Window Scheduling for LLM Inference

  SlidingServe introduces a sliding-window-based scheduler that combines a batch latency predictor, dynamic chunking, multi-level priority sorting, and DP-based batch construction to improve LLM serving capacity by up to 30% while cutting SLO violations by 16–53%.

• 06-07 中

  SlidingServe：面向LLM推理的SLO感知滑动窗口调度

  SlidingServe通过批次延迟预测器、动态分块机制、多级优先级排序和基于动态规划的批次构造，在保证SLO的同时将LLM在线推理的服务容量提升最高30%，SLO违约率降低16–53%。

• 06-06 EN

  DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

  DeepSeek-R1 demonstrates that frontier-level mathematical and code reasoning can emerge from pure reinforcement learning without any human-annotated reasoning trajectories, rivaling OpenAI o1 on AIME and Codeforces benchmarks.

• 06-06 中

  DeepSeek-R1：用强化学习激发大语言模型的推理能力

  DeepSeek-R1 证明，前沿级别的数学与代码推理能力可以从纯强化学习中涌现，无需任何人工标注的推理轨迹，在 AIME 和 Codeforces 上与 OpenAI o1 旗鼓相当。

• 06-04 EN

  Llumnix: Dynamic Scheduling for Large Language Model Serving

  Llumnix brings OS-style process rescheduling to LLM serving: it migrates in-flight requests and their KV cache across GPU instances for continuous load balancing, defragmentation, and SLO-aware priority — via a near-zero-downtime pre-copy migration mechanism and a virtual-usage abstraction, achieving up to 15x lower P99 TTFT and 36% cost savings on a 16-GPU cluster.