Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

On the Nature of Attention Sink that Shapes Decoding Strategy in MLLMs

About

Large language models and their multimodal extensions have achieved remarkable success across diverse tasks, yet the internal mechanisms that govern their reasoning behaviour remain partially understood. In particular, the attention sink, a token that attracts disproportionate attention mass, has been observed in transformer architectures, but its role is still unclear. Our goal is to understand what attention sinks represent and how they shape model behaviour during inference, rather than considering them as incidental artifacts. Through our analysis, we find that attention sink representations encode structured global information that influences the decoding process. Building on our findings, we introduce OutRo, a lightweight inference-time strategy that leverages the sink token to enhance contextual representations: (i) non-sink token representations are aligned with the sink representation in the feature space; and (ii) the sink token is allowed to attend beyond the causal constraint, facilitating information exchange with non-sink tokens. This design enhances the reasoning process without requiring additional forward passes or access to attention maps. Based on extensive experiments, OutRo consistently improves performance across representative MLLMs on seven video QA benchmarks and demonstrates strong generalisation, while incurring only a 1.1x decoding overhead.

Suho Yoo, Youngjoon Jang, Joon Son Chung• 2026

Related benchmarks

TaskDatasetResultRank
Video Question AnsweringActivityNet-QA
Accuracy44.41
376
Video Question AnsweringVideoMME
Accuracy68.33
210
Video Question AnsweringActivityNet
Accuracy49.92
22
Video Question AnsweringVideo-Holmes
Average Score46.76
12
Audio-Visual QAOmniBench
Accuracy48.25
6
Audio-Visual QAAVUT
Accuracy66.57
6
Audio-Visual QAAVHBench
Accuracy73.78
6
Audio-Visual QADailyOmni
Accuracy55.56
6
Visual-Only QAVideoHolmes
Accuracy47.63
6
Visual-Only QAVideoMME Medium
Accuracy73.89
6
Showing 10 of 10 rows

Other info

Follow for update