Asymmetric self-play has emerged as a promising paradigm for post-training large language models, where a teacher continually generates questions for a student to solve at the edge of the student's learnability. Although these methods promise open-ended data generation bootstrapped from no human data, they suffer from one major problem: not all problems that are hard to solve are interesting or informative to improve the overall capabilities of the model. Current asymmetric self-play methods are goal-agnostic with no real grounding. We propose Guided Asymmetric Self-Play (GASP), where grounding is provided by real-data goalpost questions that are identified to pose a hard exploration challenge to the model. During self-play, the teacher first generates an easier variant of a hard question, and then a harder variant of that easier question, with the goal of gradually closing the gap to the goalpost throughout training. Doing so, we improve pass@20 on LiveCodeBench (LCB) by 2.5% over unguided asymmetric self-play, and through the curriculum constructed by the teacher, we manage to solve hard goalpost questions that remain out of reach for all baselines.
@article{arxiv.2603.15957,
title = {GASP: Guided Asymmetric Self-Play For Coding LLMs},
author = {Swadesh Jana and Cansu Sancaktar and Tomáš Daniš and Georg Martius and Antonio Orvieto and Pavel Kolev},
journal= {arXiv preprint arXiv:2603.15957},
year = {2026}
}
Comments
Accepted at ICLR 2026 Workshop on AI with Recursive Self-Improvement (RSI 2026) as Spotlight, and ICLR 2026 Workshop on Lifelong Agents (LLA 2026)