Page 2 / 36
432 posts in total. Keep on posting.
Showing posts 13–24 of 432. Each entry opens locally on this site; legacy Hexo posts link back to their original article at the bottom for reference.
2026
- 中
大模型强化学习中的训推不一致:DLC 阅读笔记
从三组条件概率出发,拆解数值误差、策略滞后与采样器重同步,并核对训练稳定性和时间开销的证据。
- EN
Silent Instrument Errors | What an agent leaderboard actually measures
A technical reading of termination-gated scoring, duplicated billing spans, contaminated trajectories, and versioned replay in agent benchmarks.
- 中
Silent Instrument Errors 阅读笔记|智能体榜单的测量误差
从航空客服的漏评分出发,拆解智能体评测中的费用重计、轨迹污染和版本重放,给出判断排名是否可靠的计算方法。
- EN
FLAMIA | When the quantizer is part of the causal mask
A technical reading of reusable microscaling, hidden causal dependencies, and the trade-offs of training attention on Blackwell.
- 中
FLAMIA 阅读笔记|量化之后,因果掩码还管用吗
从缩放尺度的依赖关系出发,理解 Blackwell 上低精度注意力训练的速度、因果边界与反向数值。
- EN
Low-Rank Key Value Attention | What a shared cache can preserve
A technical reading of shared key-value projections, associative decoding, and the evidence behind LRKV's cache and quality trade-offs.
- 中
Low-Rank Key Value Attention|共享缓存与逐头残差
从共享基底和逐头残差出发,推导 LRKV 的解码公式、缓存预算与位置编码边界,并核对论文中的实验口径。
- EN
DBW: Learning the Trade-off Between SVD Pruning and Quantization
A technical review of differentiable component precision, exact integer allocation, and the distance between low-bit weights and usable inference.
- 中
DBW 阅读笔记:让 SVD 裁剪与量化精度共同分配预算
从低秩分量、整数预算与近似梯度出发,理解联合压缩的收益,以及有效位宽到真实部署之间的距离。
- 中
Mac mini + ChatGPT 可以冒名顶替我
从 EB-1A 表格和日常软件操作出发,我把资料、Loom 子 agent 与 CUA 放进 Mac mini,由 GPT 的主 agent 直接和我沟通并推进任务。
- EN
Polar: Training Native Agent Harnesses Through the Model Boundary
A technical reading of Polar covering asynchronous rollout services, token-faithful reconstruction, reward weighting, and the limits of its efficiency evidence.
- 中
Polar 阅读笔记:在原生 Agent Harness 中训练,先把轨迹说清楚
从异步 rollout、模型 API 代理与 token 轨迹重建出发,理解 Polar 的系统设计、实验收益及其条件上下文和奖励权重边界。