Bin Hu, Yingbo Xiao · 2026 · arXiv
DOI: 10.48550/arXiv.2610.31433 · open-access PDF
A compressão de cache KV para LLMs multimodais geralmente seleciona os tokens visuais para manter uma vez, no prefixo, e reutiliza essa seleção para toda a geração. Mostramos que esta seleção de um tiro é severamente limitada na geração de forma longa: na descrição detalhada da imagem, mesmo um oráculo que seleciona com a atenção da resposta não comprimida atinge apenas 0,44-0.52 ROUGE-L acordo com a saída não comprimida, porque a evidência que o modelo lê se move à medida que a descrição se desenvolve.
Machine translation (Meta M2M100); the English original follows.
KV cache compression for multimodal LLMs usually selects the visual tokens to keep once, at prefill, and reuses that selection for the whole generation. We show that this one-shot selection is severely limited on long-form generation: on detailed image description, even an oracle that selects with the attention of the uncompressed answer reaches only 0.44-0.52 ROUGE-L agreement with the uncompressed output, because the evidence the model reads moves as the description unfolds. Re-selecting whenever attention drifts out of the kept set is the natural remedy, but we find that a fixed drift thres
Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.31433, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange