机制可解释性要的不是热力图，而是因果证据。本文用 induction head、IOI 26-head 电路、activation patchi…

In The birth & death of search engine optimization , Xe suggests

This post is in some ways an antithesis of my previous postings on hard …

attention 权重容易被读成“模型关注了哪里”，但 Jain 和 Wallace（NAACL 2019）等工作证明相关不等于因果。本文区分…

I recently had a problem where part of the solution was to do a series o…

This problem in compressed sensing is an example of a derandomisation pr…

量化、蒸馏、剪枝常被当成同一件事的三种手段——”让模型变小”，但它们改变的对象、误差来源和失败模式完全不同。本文从 weight-only 与 …

There are a lot of “12 CS books every programmer must read” lists floati…

A few months ago, I was invited to contribute an article to Scholarpedia…

KV Cache 后 decode 仍是显存带宽受限的串行循环。本文给出投机解码的接受-拒绝公式与正确性证明，推导接受率如何决定加速比，用实测数…