Saved in:
Bibliographic Details
Main Authors: Jana, Swadesh, Sancaktar, Cansu, Daniš, Tomáš, Martius, Georg, Orvieto, Antonio, Kolev, Pavel
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.15957
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914399896207360
author Jana, Swadesh
Sancaktar, Cansu
Daniš, Tomáš
Martius, Georg
Orvieto, Antonio
Kolev, Pavel
author_facet Jana, Swadesh
Sancaktar, Cansu
Daniš, Tomáš
Martius, Georg
Orvieto, Antonio
Kolev, Pavel
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2603_15957
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle GASP: Guided Asymmetric Self-Play For Coding LLMs
Jana, Swadesh
Sancaktar, Cansu
Daniš, Tomáš
Martius, Georg
Orvieto, Antonio
Kolev, Pavel
Machine Learning
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.
title GASP: Guided Asymmetric Self-Play For Coding LLMs
topic Machine Learning
url https://arxiv.org/abs/2603.15957