Search papers, labs, and topics across Lattice.
This paper investigates the limitations of attention-guided visual memory eviction in stateful multimodal assistants, specifically focusing on when visual information can be safely forgotten across dialog turns. By introducing the Causal Visual Memory Audit (CVMA), the authors reveal that current attention mechanisms often misjudge the relevance of visual regions for future queries, leading to suboptimal retention of crucial information. The findings indicate that while aggregate scores may mask these issues, there exists significant potential for improving visual memory management through better understanding of fact-specific verbalization and visual dependence.
Current attention mechanisms misjudge the relevance of visual information, risking the loss of crucial context in multimodal dialog systems.
Stateful multimodal assistants encode an image once but may answer questions about it many turns later. Attention-guided visual-KV eviction assumes that evidence irrelevant now will remain dispensable, although future questions are unknown. We ask when a visual fact is actually safe to forget and introduce the Causal Visual Memory Audit (CVMA), a paired single-prefill framework that tests what later answers lose when a visual region, the whole image, or prior assistant text becomes unavailable. On VisDial and ConvBench, current attention can rank future-useful regions worse than random even though a diagnostic marginal-utility control shows substantial selection headroom. Aggregate scores hide this failure when later turns do not need vision; controlled and stock-generated histories reveal a second escape route, in which assistant-text KV replaces image KV for facts already stated but not reliably for unstated facts. In the tested stacks, safe forgetting is supported by low future visual dependence or fact-specific verbalization---not by low current attention.
