Overview
Explore how agents convey and interpret meaning through images.
How it works
Images and game context enter visual-semantic reasoning to form hypotheses and beliefs. MACH uses those beliefs and supporting memories to choose a communication strategy, generate a metaphor and refine the proposed expression. The resulting utterance and action feed the next interaction. MetaVision-DB supports the study of image meaning and communication across games.
Image and context inform beliefs, strategy and metaphor generation. Supporting memories help refine expression before the agent produces an utterance and action.
AI-assisted method illustrationSelected contributions
- Led the construction and quality review of MetaVision-DB, covering 800 images and three communication games.
- Co-developed MACH for visual-semantic reasoning and metaphor generation.
- Evaluated agent communication in the Visual Undercover game.
Research status
Third author of “Orchestrating Secrets: Visual Metaphor Generation and Reasoning in Multi-Agent Communication Games”; in revision for ARR resubmission.