Page 3 / 36
432 posts in total. Keep on posting.
Showing posts 25–36 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
VHA: Train with Virtual Heads, Deploy as Ordinary GQA
Virtual Head Attention separates a shared training parameterization from its folded inference form, improving the cache–quality trade-off while leaving important questions about optimization and measured speed.
- 中
VHA:训练时扩展虚拟头,推理时为什么还能变回普通 GQA?
从共享 query 投影、低秩头间混合到推理前的矩阵折叠,理解 VHA 怎样改善缓存与质量的权衡,以及哪些训练收益不能直接换算成部署加速。
- EN
HarnessBandit: Which Agent Interface Should Receive the Next RL Update?
A two-signal scheduler allocates agent RL updates across execution harnesses. Its gradient proxy is useful, but reward filtering, optimizer geometry and stale observations complicate the interpretation.
- 中
HarnessBandit:下一步强化学习,该让哪个智能体接口来训练?
同一个模型通过不同 harness 训练,下一步该选谁?HarnessBandit 同时观察优势对比和梯度迁移,但过滤偏差、优化器几何与过期观测让这个问题比两个分数相加更复杂。
- EN
InfraDev: When Passing the Tests Does Not Complete the Task
Jointly designing hidden probes and task specifications makes infrastructure-agent evaluation more informative, but the judge still has boundaries.
- 中
InfraDev:测试通过了,任务真的完成了吗?
从隐藏探针、接口契约和多 GPU 评测出发,理解编码智能体为何能通过测试,却没有兑现基础设施改动的承诺。
- EN
STAR-KV: Learning the Rank Is Only Half of KV Cache Compression
How learned spectral thresholds, asymmetric key/value factorization, and low-bit cache kernels interact—and what the reported accuracy and speedups actually establish.
- 中
STAR-KV:低秩 KV 缓存怎样真正加速?
从可学习谱阈值、键值不对称分解到低比特缓存,梳理 STAR-KV 的训练、推理与实验边界,区分缓存缩小、注意力加速和批量吞吐。
- EN
KV-CoRE: What Does the Rank of a KV Cache Actually Tell Us?
A technical reading of data-dependent KV-cache spectra, exact Gram accumulation, normalized effective rank, and the distance between reconstruction optimality and useful generation.
- 中
KV-CoRE:KV 缓存的秩,到底能告诉我们什么?
从数据相关的奇异值谱出发,理解 Gram 矩阵累加、归一化有效秩,以及重建最优为什么仍不能保证回答质量。
- EN
iS-KV: When the Basis Moves, the History Must Move with It
A technical reading of online KV compression through block-incremental SVD: synchronized coordinates, local optimality, reasoning quality, and the gap between smaller persistent storage and faster inference.
- 中
iS-KV:低秩基底变了,历史坐标也得一起改
从增量 SVD 的坐标一致性出发,理解 iS-KV 如何保留历史位置、近似 KV 表示,以及为什么持久缓存变小还不等于推理更快。