HackerTrans
トップ新着トレンドコメント過去質問紹介求人

thw20

no profile record

投稿

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

jeffreywong20.github.io
1 ポイント·投稿者 thw20·2 か月前·0 コメント

Towards understanding multiple attention sinks in LLMs

github.com
1 ポイント·投稿者 thw20·4 か月前·2 コメント

The Existence and Behavior of Secondary Attention Sinks

arxiv.org
1 ポイント·投稿者 thw20·5 か月前·0 コメント

コメント

thw20
·3 か月前·議論
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 か月前·議論
The up to date paper documenting and analysing the observation is now available on arxiv!
thw20
·4 か月前·議論
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 か月前·議論
This is so amazing. What a masterpiece for intro to reinforcement learning in llm.