English

AR$^2$: Adversarial Reinforcement Learning for Abstract Reasoning in Large Language Models

Computation and Language 2025-09-05 v1 Artificial Intelligence Machine Learning

Abstract

Abstraction--the ability to recognize and distill essential computational patterns from complex problem statements--is a foundational skill in computer science, critical both for human problem-solvers and coding-oriented large language models (LLMs). Despite recent advances in training LLMs for code generation using reinforcement learning (RL), most existing approaches focus primarily on superficial pattern recognition, overlooking explicit training for abstraction. In this study, we propose AR2^2 (Adversarial Reinforcement Learning for Abstract Reasoning), a novel framework explicitly designed to enhance the abstraction abilities of LLMs. AR2^2 employs a teacher model to transform kernel problems into narrative-rich, challenging descriptions without changing their fundamental logic. Simultaneously, a student coding model is trained to solve these complex narrative problems by extracting their underlying computational kernels. Experimental results demonstrate that AR2^2 substantially improves the student model's accuracy on previously unseen, challenging programming tasks, underscoring abstraction as a key skill for enhancing LLM generalization.

Keywords

Cite

@article{arxiv.2509.03537,
  title  = {AR$^2$: Adversarial Reinforcement Learning for Abstract Reasoning in Large Language Models},
  author = {Cheng-Kai Yeh and Hsing-Wang Lee and Chung-Hung Kuo and Hen-Hsen Huang},
  journal= {arXiv preprint arXiv:2509.03537},
  year   = {2025}
}

Comments

7 pages, accepted by CIKM 2025 as a short paper

R2 v1 2026-07-01T05:19:41.513Z