HackerTrans
TopNewTrendsCommentsPastAskShowJobs

thw20

no profile record

Submissions

Simple, zero overhead way to compress model, KV cache via Low-Rank Decomposition

jeffreywong20.github.io
1 points·by thw20·2 เดือนที่ผ่านมา·0 comments

Towards understanding multiple attention sinks in LLMs

github.com
1 points·by thw20·4 เดือนที่ผ่านมา·2 comments

The Existence and Behavior of Secondary Attention Sinks

arxiv.org
1 points·by thw20·5 เดือนที่ผ่านมา·0 comments

comments

thw20
·3 เดือนที่ผ่านมา·discuss
Good work! This is very interesting. Here's a related work that construct low-rank approximation for attention: https://arxiv.org/abs/2505.12942.

Maybe the idea of Query calibration matrix Rxx is of interest to the author!
thw20
·4 เดือนที่ผ่านมา·discuss
The up to date paper documenting and analysing the observation is now available on arxiv!
thw20
·4 เดือนที่ผ่านมา·discuss
This project reveals an interesting phenomena, where LLM converts semantic non-informative tokens to attention sinks through middle layer MLP.

The converted sinks are termed secondary attention sinks as they are weaker then BOS attention sinks.

This might be related to layer specialisation in LLM!
thw20
·4 เดือนที่ผ่านมา·discuss
This is so amazing. What a masterpiece for intro to reinforcement learning in llm.