AbstRaL: Teaching LLMs Abstract Reasoning via Reinforcement to Boost Robustness on GSM Benchmarks

Recent research highlights that smaller large language models (LLMs) exhibit significant weaknesses in robust reasoning, particularly in out-of-distribution (OOD) scenarios where slight alterations to familiar questionssuch as changing names, numbers, or adding distractionslead to substantial drops in accuracy. To address this, the study introduces AbstRaL, a reinforcement learning-based approach that trains LLMs to focus on the underlying logic of reasoning problems by generating synthetic variations, thereby enhancing their ability to generalize beyond surface-level cues. This development aims to improve the reliability and generality of LLMs across logic, mathematics,