MATE builds joint attention into agents for zero-shot teaming
TL;DR
- MATE is a multi-agent RL method that aligns agents' visual attention on scene-salient objects instead of partner-specific conventions.
- The authors test MATE on three benchmarks: a Card Alignment Game, Level-Based Foraging, and OvercookedV2.
- The abstract reports MATE 'consistently' improves coordination with unknown partners but publishes no per-benchmark numbers.
A new paper on arXiv proposes building human-style joint attention into reinforcement learning agents so they can coordinate with partners they have never trained alongside. The authors, Giulia Benintendi, Constantin Ruhdorfer, Fabian Kögel and Andreas Bulling, call the method MATE, short for Mutual Attention for zero-shot TEaming.
"Joint attention - the human ability to share a common visual or cognitive focus with others - enables a meeting of minds that lets us coordinate even with unfamiliar partners," the abstract reads. MATE pushes agents to align their visual attention on "scene-salient objects" rather than lean on "arbitrary partner-dependent conventions established during training." The authors frame this as a step past symmetry-breaking approaches, which they say "merely prevent brittle conventions from emerging."
The method is evaluated on three benchmarks: a purpose-built Card Alignment Game, Level-Based Foraging, and OvercookedV2. The paper reports that experiments "consistently show that a joint-attention-inspired signal improves coordination with unknown partners." The abstract publishes no per-benchmark numbers.
Shared on Bluesky by 1 AI expert
-
Eugene Vinitsky @eugenevinitsky.bsky.social: Constantin Ruhdorfer is really on a run of quite interesting papers: arxiv.org/abs/2610.06025 →
Originally reported by arxiv.org
Read the original article →Original headline: Grounded Joint-Attention Other-Play for Zero-Shot Coordination