In the paper "Rewarding Chatbots for Real-World Engagement with Millions of Users" (https://doi.org/10.48550/arXiv.2303.06135) from March of this year, researchers were able to increase the user retention of chatbots by 30% using a simple reinforcement learning routine and training a reward model against proxy 'pseudo-labels' (conversation time, swipes, user ratings for responses).

I would be interested in seeing how much this kind of capacity scales. In principle, reinforcement learning should be really good at optimizing for this kind of goal, but maybe there's something I'm missing.

First poll I'm making on manifold, so there are probably some conventions I'm not aware of when it comes to framing questions in a way that allows them to be cleanly resolvable, but as of now my plan is that in 2024 I'll see what further results have been published on this and pick the best result that an academic paper has been able to show.