Parametric Open Source Games

The paper's core contribution is identifying a precise learning rate threshold where gradient descent flips from defection to cooperation in symmetric games. This isn't just 'cooperation emerges sometimes' - it's a quantified phase transition that formalizes when self-interest produces prosocial behavior, which prior multi-agent RL work treated as an empirical accident rather than a predictable phenomenon.

This connects directly to the Heavy-Ball Q-Learning paper from the same day, which also formalizes convergence guarantees and identifies conditions where acceleration techniques outpace standard methods. Both papers move multi-agent learning from 'it works empirically' to 'here's when and why it works.' The parametric games work is more foundational (game theory layer), while the Q-learning paper is more applied (RL layer), but they share the same intellectual project: replacing intuition with formal conditions. The Bayesian inference framework from earlier today also echoes this theme of moving beyond global metrics to local, parameter-dependent analysis.

If researchers successfully instantiate the identified cooperation threshold in a trained neural network policy on a standard benchmark (e.g., iterated prisoner's dilemma or a multi-agent RL environment), and the threshold matches the theoretical prediction within a factor of 2, that confirms the formalism translates to practice. If the threshold only holds for toy symmetric games and breaks under asymmetry or realistic reward structures, the contribution remains primarily theoretical.