PopuLoRA: Co-Evolving LLM Populations for Reasoning Self-Play
The paper introduces PopuLoRA, a novel framework for reinforcement learning with large language models (LLMs). It utilizes a population-based approach where specialized teachers and students engage in self-play to enhance problem-solving capabilities. The results demonstrate that this method outperforms traditional single-agent training on various benchmarks despite lower training-time rewards.
- ▪PopuLoRA employs a population-based asymmetric self-play framework for reinforcement learning with verifiable rewards.
- ▪Teachers propose problems while students solve them, facilitating a co-evolutionary process that enhances problem complexity over time.
- ▪The population mean performance surpasses a single-agent baseline on multiple code and math benchmarks.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.16727 (cs) [Submitted on 16 May 2026] Title:PopuLoRA: Co-Evolving LLM Populations for Reasoning Self-Play Authors:Roger Creus Castanyer, Geoffrey Bradway, Lorenz Wolf, Maxwill Lin, Augustine N. Mavor-Parker, Matthew James Sargent View a PDF of the paper titled PopuLoRA: Co-Evolving LLM Populations for Reasoning Self-Play, by Roger Creus Castanyer and 5 other authors View PDF HTML (experimental) Abstract:We introduce PopuLoRA, a population-based asymmetric self-play framework for reinforcement learning with verifiable rewards (RLVR) post-training of LLMs.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
