Decentralized Learning Enables Emergent Communication Between Heterogeneous Visual Agents
A recent study published on arXiv (2605.11695) examines the development of shared symbols among diverse visual agents through decentralized learning, even in the absence of a unified communicative goal. In the Metropolis-Hastings Captioning Game (MHCG), two agents interact by exchanging discrete token sequences, refining their models based on local perceptual data, while a listener evaluates proposals using an MH-style criterion. The research delves into the possibility of common symbols arising without shared perceptual access and how the differences in private visual spaces influence the language that develops. This investigation sheds light on a less-explored facet of emergent communication, particularly in decentralized environments with varying visual representations among agents.
Key facts
- arXiv paper 2605.11695 studies emergent communication between heterogeneous visual agents
- Agents use decentralized learning without a shared external communicative objective
- Metropolis-Hastings Captioning Game (MHCG) is the experimental setting
- Agents exchange discrete token sequences and update models via local perceptual evidence
- Listener accepts or rejects proposals using an MH-style criterion
- Research asks whether common symbols can arise without shared perceptual access
- Similarity between private visual spaces constrains language content and symmetry
- Focuses on an underexplored aspect of emergent communication
Entities
Institutions
- arXiv