The first survey on Attention Sink in Transformers structures the literature around fundamental utilization, mechanistic interpretation, and strategic mitigation.
Active-dormant attention heads: Mechanistically demystifying extreme-token phenomena in llms
5 Pith papers cite this work. Polarity classification is still indexing.
5
Pith papers citing it
citation-role summary
background 2
citation-polarity summary
roles
background 2representative citing papers
OScaR mitigates token norm imbalance via canalized rotation and omni-token scaling to enable near-lossless INT2 KV cache quantization with up to 3x decoding speedup and 5.3x memory reduction.
Attention sinks emerge in language models from softmax-induced token dependence on attention scores and do not appear when using sigmoid attention without normalization in models up to 1B parameters.
LongAct uses saliency from high-magnitude activations to guide sparse weight updates in long-context RL, yielding about 8% gains on LongBench v2 across multiple algorithms.
citing papers explorer
-
Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation
The first survey on Attention Sink in Transformers structures the literature around fundamental utilization, mechanistic interpretation, and strategic mitigation.
-
OScaR: The Occam's Razor for Extreme KV Cache Quantization in LLMs and Beyond
OScaR mitigates token norm imbalance via canalized rotation and omni-token scaling to enable near-lossless INT2 KV cache quantization with up to 3x decoding speedup and 5.3x memory reduction.
-
When Attention Sink Emerges in Language Models: An Empirical View
Attention sinks emerge in language models from softmax-induced token dependence on attention scores and do not appear when using sigmoid attention without normalization in models up to 1B parameters.
-
LongAct: Harnessing Intrinsic Activation Patterns for Long-Context Reinforcement Learning
LongAct uses saliency from high-magnitude activations to guide sparse weight updates in long-context RL, yielding about 8% gains on LongBench v2 across multiple algorithms.
- Attention Sink Forges Native MoE in Attention Layers: Sink-Aware Training to Address Head Collapse