PopuLoRA: Co-Evolving LLM Populations for Reasoning Self- Play

PopuLoRA introduces a population-based framework for training large language models through reinforcement learning with verifiable rewards (RLVR), using co-evolving populations of teacher and student models rather than single-agent self-play. The key innovation addresses a critical failure mode where single-agent self-play collapses into generating overly simple tasks that the model can easily solve, by having separate teacher models generate tasks and student models solve them, forcing teachers to continually produce harder and more diverse problems as students improve. This adaptive curriculum approach enables models to develop more sophisticated reasoning capabilities than traditional self-play or fixed task distributions allow.