Why Self-Play Works for Go but Struggles with LLMs

i_dg23 · x · 2026-08-01

The author explores why self-play excels in two-player zero-sum (2p0s) games like Go, Poker, and StarCraft, but is much harder to apply in real-world domains.

The core reason is that self-play converges to a minimax equilibrium in 2p0s games, which is an unbeatable strategy in expectation. However, this isn't always optimal: against weak opponents, playing minimax might not maximize payoffs (e.g., failing to extract maximum profit from bad poker players). Because the real world is messy and non-zero-sum, self-play hasn't seen many successes in LLMs yet.

Original post →

More from Research

Research channel →