Resolution as Attention: arXiv paper renders agent memories as images to cut token use
A new arXiv paper proposes treating visual resolution as attention, letting LLM agents store past interactions as images whose detail level matches each memory's relevance.
In plain words
AI agents that work on long tasks keep records of everything they have done, and those records get expensive to process. The paper's idea is to turn each past interaction into a picture and store important ones in high detail while keeping less important ones blurry. A small trained model decides how sharp each memory should be.
What was announced
The authors argue that memory research has focused on what information agents keep or retrieve, while paying little attention to how much perceptual detail each stored interaction deserves. Their mechanism, Resolution as Attention (RAA), renders past interactions as images and adjusts each image's resolution at every step according to its relevance to the agent's next decision. Critical interactions keep fine-grained detail, while less relevant ones consume fewer visual tokens.
To set the fidelity of each memory, the team trained a lightweight relevance estimator with guidance from a teacher model, then optimized it to preserve the agent's decision consistency while lowering visual-token cost. In experiments across multiple tasks and backbones, the authors report token consumption fell by more than 50% without hurting task performance. They add that RAA works as a plug-in layer alongside existing memory management methods, which they say produced further savings and suggests adaptive perception is a complementary dimension of agent memory.
What it does not show
- The evidence is the paper's abstract only; the reported token savings and performance results have not been independently verified, and the full paper's methods and task details are not covered here.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.