PopuLoRA co-evolves LLM populations for reasoning self-play

Arxiv CS.AI (top papers)

Introduces PopuLoRA, a population-based asymmetric self-play framework for RLVR post-training. Teachers and students are specialized LoRA adapters on a shared frozen base: teachers propose verifiable tasks, students try to solve them. As students improve, teachers must search for harder tasks, creating an adaptive curriculum.

populoraLLMReinforcement learning

See 1 more source

PopuLoRA: Co-Evolving LLM Populations for Reasoning Self- Play27 days agoAMavorParkerDiscuss

Arxiv CS.AI (top papers)·Roger Creus CastanyerGeoffrey BradwayLorenz WolfMaxwill LinAugustine N. Mavor-ParkerMatthew James Sargent

28 days ago