This week's arXiv batch clusters around a common interest in making vision-language systems more structured and cooperative. Two papers explicitly target multi-agent settings. The first, Collaborative Memory for Multi-Agent VLM Systems, proposes that agents inspecting different image regions or video frames can benefit from a shared memory. The second, MACE: Memory-Agent Co-Evolution, applies a similar idea to LLM-based agents, using adaptive memory graphs to store and reuse collaboration traces such as planning steps, verification results, and failure repairs. Both assume that remembering how agents work together is as important as the visual processing itself.
A third paper, VeriFuse, takes a more cautious stance. It notes that while vision-language models are strong at scene understanding, asking them to directly regress 3D coordinates in cooperative perception is problematic. Instead, the authors propose bounded arbitration and reason-guided refinement, meaning the VLM contributes high-level judgments while dedicated modules handle precise geometry. This contrasts with the memory-centric approaches, which focus on storing procedures rather than constraining the VLM's role.
The remaining two papers are tangential to the multi-agent thread. TAPe+ML introduces a compact structured representation for multi-task vision, encoding relations among perceptual elements before recognition rather than operating on raw pixels. Extending Decoupled Attention addresses multi-channel imaging, where each channel is a distinct semantic signal, and adapts vision encoders via masked training. Neither discusses agents or memory, so the connection to the first three is thematic at best—both are about improving perception through explicit structure, but they operate at a different level of the stack. Readers should treat these as separate research directions rather than parts of a single movement.}}{