Page 5 / 36
432 posts in total. Keep on posting.
Showing posts 49–60 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
SpecScale: Making Speculative Reasoning Search Work Under Serving Load
Why early reasoning branches can reduce idle time yet slow a busy GPU, and how pruning, shared-prefix computation, and batched verification change that trade-off.
- 中
SpecScale:推理分支提前展开后,怎样避免越跑越慢?
从提前剪枝、相同前缀的计算去重到延迟批量验证,理解推测搜索为什么能减少等待,却也可能拖慢繁忙的推理服务。
- EN
SpecReason: When Can a Small Model Take the Next Reasoning Step?
A close reading of step-level speculative reasoning, its verification economics, and the gap between preserving answer quality and preserving a model distribution.
- 中
SpecReason:哪些推理步骤可以放心交给小模型?
从逐步验证的成本、语义接受的边界和实验预算出发,理解小模型如何加速长链推理,以及为什么更快不等于无损。
- EN
EvoRubric: Who Teaches a Model to Write Its Own Grading Criteria?
A close reading of shared-policy rubric learning, the limits of peer consensus, and the evidence behind open-ended reinforcement learning.
- 中
EvoRubric:让模型自己写评分标准,谁来判断标准对不对?
从带符号评分、留一共识到动态记忆,理解开放式强化学习怎样学习评价标准,以及共识为什么仍不等于正确。
- EN
STOP: When the Code Improver Becomes Its Own Optimization Target
A technical reading of recursive scaffold improvement: what STOP changes, why the evaluator becomes the bottleneck, and what its experiments establish.
- 中
STOP:让代码改进器改写自己,怎样判断它真的变好了?
从两层优化、递归改写和迁移实验出发,拆解 STOP 的方法、统计假设、计算成本,以及较弱模型为何越改越差。
- EN
ADAS: What Changes When Agents Design Other Agents?
A close reading of code-space agent search, archive-driven discovery, transfer results, and the evaluation needed to separate better workflows from better selection.
- 中
ADAS:让智能体设计智能体,究竟改变了什么?
从程序搜索、历史档案和跨任务迁移出发,理解自动设计智能体的收益,以及它与改进设计者自身之间的距离。
- EN
Dobi-SVD: What Changes When We Compress Activations Before We Compress Weights?
A technical reading of differentiable rank selection, shared activation subspaces, and the rank–precision tradeoff behind Dobi-SVD.
- 中
Dobi-SVD:先截断激活,再压缩权重,中间还差什么?
从可微秩分配、共享子空间和混合精度存储出发,理解 Dobi-SVD 的有效之处,以及局部最优与实际部署之间的距离。