AlphaEvolve finds new multiagent algorithms

This is more important than another agent demo because AlphaEvolve is being used to generate new learning rules in a domain where progress usually comes from slow human iteration. If results like this keep transferring to other research areas, algorithm design itself becomes a more automated part of the ML stack.

• –VAD-CFR and SHOR-PSRO attack two established families of game-theoretic learning methods instead of proposing a completely new toy setup
• –The paper’s strongest signal is that AlphaEvolve found non-intuitive changes humans did not manually engineer, which is the real promise of automated discovery
• –Imperfect-information games are a serious testbed, so beating strong baselines here is more meaningful than a flashy but narrow coding benchmark
• –The main caveat is external validity: these gains still need to generalize beyond game settings before this becomes a broader research workflow shift