Unverified50% confidenceFactExact time
In multimodal large models, an image is split by the vision encoder into a large number of patch tokens, and the attention distribution is long-tailed: a few visual tokens receive high attention while a large number receive very low attention.
1
Sources
50%
Confidence
Long-term
Relevance
11/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn multimodal models, reference is usually realized through attention alignment between text tokens and visual tokens.80% similarUnverifiedToken quota consumption for multimodal Agent tasks is often far higher than for pure text conversations, because image inputs must be encoded as visual tokens, each Computer Use screenshot introduces new visual input, and long multi-step reasoning chains accumulate intermediate reasoning tokens.75% similarUnverifiedIn multimodal agents, visual reasoning refers to the mechanism by which a model converts a page screenshot into image tokens and then parses it frame by frame to determine the positions of clickable elements; a high-resolution screenshot often consumes hundreds to thousands of image tokens, and inference must be redone every time the page state changes.73% similarUnverifiedThe current mainstream multimodal understanding approach is to align image features output by a visual encoder into the language model's embedding space via a projection layer or cross-attention, so that text tokens and image tokens participate in self-attention computation within the same sequence (early fusion).72% similarUnverified大模型中存在注意力汇聚（Attention Sink）现象，即序列最开头的几个 token 往往保持较高的注意力权重72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1137657API
curl https://kongchang.com/api/v1/knowledge/claims/1137657MCP
get_claim(id=1137657)