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

Abstract

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.

Figures (14)

• Figure 1: VLM criterion $\Phi_{\text{VLM}}$ incorrectly identifies many semantic tokens as sinks.(a) Number of sink tokens across layers under $\Phi_{\text{VLM}}$ and $\Phi_{\text{LLM}}$. $\Phi_{\text{VLM}}$ rapidly over-identifies sink tokens in deeper layers, unlike the sparse behaviour of $\Phi_{\text{LLM}}$. (b) Qualitative visualisation for the queries "Is the helicopter visible in the video?". Although the queried object tokens receive high attention, indicating that the model appropriately focuses on them, they are nevertheless identified as sinks under $\Phi_{\text{VLM}}$.
• Figure 2: Head pruning on AVHBench. Sink scores pruning performs better than random pruning.
• Figure 3: High sink attention does not imply head redundancy.(a) Performance change after removing individual heads, distinguishing improvement and decrease heads. (b) High sink scores are observed in both improvement and decrease heads, indicating that sink-heavy heads are not always redundant.
• Figure 4: Geometric analysis of sink token representations.(a) Sink keys and values exhibit substantially smaller norms than other tokens. (b) Their activations are not uniformly small across dimensions. (c) In t-SNE, sink keys form a coherent cluster near the query distribution. (d) Despite their small norms, the cosine similarity between queries and sink keys remains high, resulting in large attention scores
• Figure 5: Sink token attention score. Compared to the baseline, zeroing the top-1 dominant dimension in the sink key representations substantially reduces attention to sink tokens across layers.