IC-945LLaMA-2, MPT, Falcon, Pythia, and BERT-base-uncased allocate disproportionate attention to initial tokens regardless of their semantic content

Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, Mike Lewis

SourceEfficient Streaming Language Models with Attention Sinks

The paper visualizes attention maps across all layers and heads of LLaMA-2-7B and LLaMA-2-70B, finding that beyond the bottom two layers the model heavily attends to the initial token. A substitution experiment on LLaMA-2-13B shows that replacing the first four tokens with linebreak characters still restores perplexity to near-normal levels (5.60 vs 5.40), confirming the effect is positional rather than semantic. Across MPT-7B, Falcon-7B, Pythia-12B, and LLaMA-2-7B, introducing four initial tokens alongside a recent window restores perplexity from catastrophic levels to near-normal. BERT-base-uncased shows the same pattern with its [SEP] token receiving disproportionately high attention in most layers. The authors attribute the phenomenon to the softmax constraint requiring attention scores to sum to one, forcing the model to dump excess attention onto globally visible initial tokens.

Evidence
correlational
Key metric
LLaMA-2-13B: 0+1024 ppl 5158.07, 4+1020 ppl 5.40, 4"\n"+1020 ppl 5.60; LLaMA-2-7B: 0+4096 ppl 3359.95, 4+4092 ppl 9.59; MPT-7B: 0+2048 ppl 460.29, 4+2044 ppl 14.99; Falcon-7B: 0+2048 ppl 17.90, 4+2044 ppl 12.12; Pythia-12B: 0+2048 ppl 21.62, 4+2044 ppl 12.09
Caveat
The attention sink observation is directly visualized on LLaMA-2-7B and LLaMA-2-70B; other model families are supported by perplexity recovery experiments rather than direct attention map visualization. The BERT result is a single attention map on one sentence.
Model
Llama 2 / Llama 2 base Llama 2 7B, Llama 2 13B, Llama 2 70B, MPT MPT-7B, Falcon Falcon-7B, Pythia Pythia-12B, BERT BERT-base-uncased
Concepts
Positional bias
Datasets
PG19 [eval], ARC [eval]
Methods
StreamingLLM [primary]
Related findings
IC-946
Extraction
automatic-extraction