Page 4 / 36
432 posts in total. Keep on posting.
Showing posts 37–48 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
HyScale: Replaying Sparse Selections
Why learned sparse attention still pays for dense selection during training, and how exact on-the-fly top-k, selection replay, and dependency-aware scheduling reduce that cost.
- 中
HyScale:稀疏注意力训练,为什么还在重复做稠密选择?
从完整评分表到可重放的索引,理解 HyScale 如何减少稀疏注意力训练中的选择、重算和通信开销,以及精确性与加速结论的边界。
- EN
Self-Indexing Attention: One Sign Field for Retrieval and Compression
How transformed key signs become both a sparse-attention index and part of a compressed cache, and why this elegant representation still needs careful accounting of retrieval error, memory, and end-to-end speed.
- 中
Self-Indexing Attention:一组符号位,怎样同时服务检索与压缩?
从共享旋转、符号检索和组内池化出发,逐步理解自索引注意力的表示设计,并区分缓存字节数、算子加速、整模型收益与任务级质量损失。
- EN
Self-HPO: When the Harness and the Policy Learn Together
A close reading of alternating harness search and policy reinforcement learning, with derivations, crossed-checkpoint evidence, and the boundary between solving a task and delivering its patch.
- 中
Self-HPO:让模型和运行框架一起进化,收益来自哪里?
精读 harness 搜索与策略强化学习的交替优化:从二值奖励的学习信号,到交叉评测、补丁交付失败与策略对接口的依赖。
- EN
RRSI: When Should a Self-Evolving Agent Keep Its Changes?
A close reading of regularized harness evolution: adaptive overfitting, sparse edits, conservative selection, transfer evidence, and the limits hidden in the appendices.
- 中
RRSI:Agent 改进自己之后,哪些修改值得留下?
从搜索集过拟合出发,理解 harness 自我进化中的稀疏编辑、保守选择与迁移证据,并细读成本、统计区间及附录中的规则边界。
- EN
VERSE: What Makes a Self-Evolving Harness Optimizer Worth Trusting?
A close reading of VERSE: executable evidence, two coupled harnesses, adaptive validation, and the limits of self-improvement across tasks and models.
- 中
VERSE:让优化器改进自己,怎样才不只是越改越自信?
精读 VERSE 的执行验证、双层 harness 优化、验证集复用与跨模型迁移,分析自我进化何时有依据、何时会把噪声写成经验。
- EN
Harness²: Can an agent improve its workflow without a correctness oracle?
A close reading of task-local harness search, cross-task improvement experience, and the gap between observing better behavior and verifying a correct answer.
- 中
Harness²:没有标准答案,agent 怎样改进自己的工作方式?
从单任务内的提议、探测和组合,到跨任务保留经验,拆解 Harness² 的两层递归,以及行为改善与正确性之间仍然存在的距离。