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

PopuLoRA introduces a population-based self-play framework for RLVR post-training where teacher and student LoRA adapters on a shared frozen base co-evolve via weight-space mutation and crossover operators. Cross-population evaluation addresses the self-calibration collapse common in single-agent self-play. Teams doing RLVR fine-tuning for reasoning tasks should watch this as an alternative to GRPO or self-play baselines, though benchmark numbers are absent from the excerpt.