Pinned
People adapt their language to communicate more efficiently over time. How can we make models do this? In our recent work, we trained models in self-play, and found that using the right incentives can make models adapt to communicate efficiently even without human demonstrations.
