331 posts in total. Keep on posting.

2026

• 07-13 中

  Long-Horizon-Terminal-Bench 阅读笔记：密集奖励评估揭露 Agent 长程执行瓶颈

  LHTB 用 46 个容器化终端任务替代二元通过/失败评分，以子任务级密集奖励量化 Agent 进度——结果发现即使最强模型 GPT-5.5 也只有 15.2% 通过率，79% 的失败源于超时而非单步推理错误。

• 07-11 EN

  DPPO: Fixing PPO's Ratio Clipping by Constraining Divergence Instead of a Single Token's Probability

  DPPO shows that PPO's ratio-clipping trust region is a noisy single-token proxy that over-penalizes rare tokens and under-penalizes common ones — and replaces it with a cheap, direct estimate of policy divergence that stabilizes and speeds up LLM RL fine-tuning.

• 07-11 中

  DPPO 阅读笔记：与其信任一个采样 token 的比率，不如直接约束策略散度

  DPPO 指出 PPO 的比率裁剪信任域本质上是对单个采样 token 的高噪声估计——它会过度惩罚罕见 token、又对高频 token 惩罚不足；论文用一个几乎零开销的散度直接估计取而代之，显著提升了大模型强化学习的稳定性与效率。

• 07-10 EN

  FlashSVD v1.5: Why Low-Rank LLMs Don't Get Faster on Their Own

  FlashSVD v1.5 shows that SVD-compressed transformers stay slow at serving time not because low-rank arithmetic is weak, but because a naive runtime shatters every layer into hundreds of tiny kernel launches — and that a unified dense-KV, packed-MLP, per-layer-graph-replay runtime turns the same fixed checkpoints into up to 2.55x faster decoding.

• 07-10 中

  FlashSVD v1.5：为什么低秩大模型不会自动变快

  FlashSVD v1.5 揭示了 SVD 压缩后的 Transformer 推理依然缓慢的根本原因不是低秩运算本身效率低，而是朴素运行时将每层拆散成数百次细碎内核调用；论文提出密集 KV 解码注意力、打包 MLP 投影与逐层 CUDA 图重放三项机制，在不改变任何压缩检查点的前提下将解码速度提升最高 2.55 倍。

• 07-09 EN

  RATrain: Training-State Lifecycle Scheduling for Dense LLM Training on Bandwidth-Constrained Heterogeneous Supercomputers

  RATrain reframes dense LLM pipeline training as a fine-grained training-state lifecycle scheduling problem, letting a bandwidth-starved, memory-constrained domestic HPC accelerator match 8xA800-class throughput without changing training semantics.

• 07-09 中

  RATrain：面向带宽受限异构超算的稠密大模型训练状态生命周期调度

  RATrain 把稠密大模型的流水线训练重新建模为细粒度的训练状态生命周期调度问题，让一台算力尚可但显存和带宽都很紧张的国产超算加速器，在不改变训练语义的前提下逼近 8xA800 级别的训练吞吐。

• 07-08 EN

  DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

  DSpark pairs a semi-autoregressive drafter — a parallel backbone plus a lightweight sequential head — with a hardware-aware confidence scheduler that casts verification-length selection as a provably lossless throughput-maximization problem, shipping in DeepSeek-V4 production for a 60-85pct per-user speedup.

• 07-08 中

  DSpark:半自回归生成与置信度调度的投机解码

  DSpark 用「并行主干 + 轻量顺序修正头」的半自回归草稿模型，配合一个把验证长度选择问题严格建模为可证明无损、感知硬件负载的吞吐最大化问题的置信度调度器，已在 DeepSeek-V4 生产环境中实现单用户 60-85% 的生成加速。

• 07-07 EN

  RSPO: Reward-Swap Policy Optimization for Multi-Turn LLM Agents

  RSPO bridges the gap between sparse outcome rewards and dense process rewards in multi-turn LLM agent training by using a reward-swap cyclic framework that expands trajectory diversity without reward misalignment.

• 07-07 中

  RSPO：多轮LLM智能体的奖励交换策略优化

  RSPO 通过奖励交换机制，让稠密过程奖励训练的探索智能体扩展轨迹多样性，再用结果奖励训练最终策略，同时避免奖励错位与奖励黑客问题。

• 07-06 EN

  The Mirage of Optimizing Training Policies: Monotonic Inference Policy Improvement for LLM RL

  MIPU exposes an overlooked objective-level flaw in LLM RL: training-side improvement does not guarantee inference-side improvement under training-inference mismatch, and proposes a two-step framework to fix it.