thw20

3 tháng trướcdiscuss

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 tháng trướcdiscuss

The up to date paper documenting and analysing the observation is now available on arxiv!

thw20

4 tháng trướcdiscuss

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 tháng trướcdiscuss

This is so amazing. What a masterpiece for intro to reinforcement learning in llm.

Submissions

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

Towards understanding multiple attention sinks in LLMs

The Existence and Behavior of Secondary Attention Sinks

comments