Mechanism-Guided Selective Unlearning for RLVR-Induced Reasoning

A team of researchers has introduced MAST, a mechanism-guided method designed to remove reasoning behaviors acquired through reinforcement learning from verifiable rewards while limiting damage to a model’s other capabilities, according to a preprint posted to arXiv on June 17 [1]. The approach, formally called Mechanism-Aligned Selective Targeting, targets a specific problem in large language models: unlearning reasoning patterns induced by RLVR without degrading performance on unrelated tasks [1]. The researchers examined matched checkpoints from supervised fine-tuning and RLVR stages on two models, Qwen2.5-Math-1.5B and Qwen3-1.7B-Base [1][2]. They found that the shift in token-level delta-log-probability between the SFT and RLVR stages differed sharply from the original SFT update, and that standard full-parameter gradient ascent caused forgetting only by damaging performance on retained portions of the MATH dataset and on GSM8K [1][2]. MAST works by ranking attention-projection tensors using three criteria: off-principal energy, update magnitude, and forget-gradient coupling magnitude. It then updates only the top-ranked subset of those tensors [1][2]. On the primary model, this selective update produced statistically significant forgetting of the targeted material, reducing MATH forget examples from 45 out of 150 to 37 out of 150, with a McNemar p-value of 0.0078 [1][2]. At the same time, GSM8K performance improved by 0.8 percentage points and MATH retain performance dipped by just 0.5 percentage points [1][2]. The advantage held across multiple random seeds, two unlearning objectives—NPO and SimNPO—and the Qwen3 architecture. In the Qwen3 setting, full-parameter unlearning caused GSM8K performance to collapse, while MAST preserved it [1][2]. The preprint, hosted on arXiv, an open-access repository that has served the physics, mathematics, and computer science communities since 1991 and now receives roughly 24,000 submissions per month, has not yet undergone peer review [6].