Page 6 / 36
432 posts in total. Keep on posting.
Showing posts 61–72 of 432. Each entry opens locally on this site; legacy Hexo posts link back to their original article at the bottom for reference.
2026
- EN
M*: What a Multimodal Serving Runtime Must Know About the Model
A technical reading of Walk Graphs, streaming policies, placement, and the conditions behind M*'s multimodal serving results.
- 中
M*:多模态模型的推理,应该怎样交给运行时?
从 Walk Graph 的依赖、循环和流式分块出发,理解 M* 的运行时设计、加速来源,以及实验尚未回答的问题。
- EN
DART: Parallel Drafts, Local Continuity, and the Real Cost of Speculation
A technical reading of DART, from prefix-conditioned parallel logits and weighted distillation to n-gram tree pruning, exact verification, and system-level limits.
- 中
DART:把草稿并行生成之后,还要补上哪些依赖?
从前缀隔离训练、位置加权蒸馏到 n-gram 树剪枝,理解 DART 的速度来源、精确验证条件和主机内存代价。
- EN
TTT-Discover: Learning During Search for One Exceptional Solution
A technical review of test-time reinforcement learning, adaptive entropic advantages, PUCT-inspired state reuse, and the evidence separating benchmark improvements from general scientific discovery.
- 中
TTT-Discover:一边搜索一边学习,为了找到一个真正出色的解
从自适应熵目标、候选状态复用到数学、GPU 算子和生物数据实验,理解测试时强化学习为何有用,以及它尚未证明什么。
- EN
AFlow: Searching for Better Agent Workflows, and Measuring What Improves
A technical reading of AFlow's program-level search, mixed parent selection, feedback memory, six-benchmark evidence, transfer, and the limits of its cost and convergence claims.
- 中
AFlow 阅读笔记:让模型搜索工作流,也把改进的代价算清楚
从两层节点、混合采样和局部经验出发,理解 AFlow 如何搜索可复用的 Agent 工作流,并重新核对性能提升、模型迁移、成本与收敛主张的边界。
- EN
AutoHarness: Learning the Rules Around an LLM Agent
A technical review of code-harness synthesis, legality-aware search, and the evidence separating reliable actions from strong game play.
- 中
AutoHarness:让智能体写出自己的规则护栏
从动作合法性到完整代码策略,理解 AutoHarness 的程序搜索、实验收益,以及测试中零错误与实际可靠性的距离。
- EN
DFlash: Why a Diffusion Drafter Can Accelerate an Autoregressive Model
A close reading of target-feature conditioning, one-step parallel drafting, exact verification, and the deployment boundaries of DFlash.
- 中
DFlash 阅读笔记:并行起草、目标特征与精确验证
从前缀接受概率和精确采样出发,理解 DFlash 的目标特征注入、单步并行草稿,以及不同并发与上下文条件下的加速边界。