Research
Resolution as Attention for Efficient Agent Memory Perception
Researchers propose Resolution as Attention (RAA), a memory perception mechanism that renders past agent interactions as images and dynamically adjusts their resolution based on relevance to the next decision. A lightweight relevance estimator, trained with teacher-model guidance and optimized for decision consistency, determines how much visual detail each interaction receives. Experiments across multiple tasks and backbones reportedly reduce token consumption by over 50% without compromising task performance. RAA can also serve as a plug-in perception layer alongside existing memory management methods.
LLM agentsagent memorymultimodal perception