mlsanityData to 14 Oct 2026

Mechanism-Guided Selective Unlearning for RLVR-Induced Reasoning

arXiv 2606.19222 ↗First posted 17 Jun 2026cs.LG cs.AIcited by 0

On the map: LLM safety and security → Machine unlearning

Abstract (arXiv)

We propose MAST (Mechanism-Aligned Selective Targeting), a mechanism-guided method for unlearning RLVR-induced reasoning with substantially lower collateral damage than standard full-parameter updates. In matched SFT/RLVR checkpoints on Qwen2.5-Math-1.5B and Qwen3-1.7B-Base, the SFT-to-RLVR increment differs sharply from the SFT update in token-level delta-log-probability, and full-parameter gradient ascent forgets only by damaging retain MATH and GSM8K. MAST ranks attention-projection tensors by off-principal energy, update magnitude, and forget-gradient coupling magnitude, then updates only the top-ranked subset. On the primary model, MAST induces statistically significant target forgetting (MATH forget 45/150 to 37/150; McNemar p=0.0078) while preserving GSM8K (+0.8 pp) and MATH retain (-0.5 pp). The advantage reproduces across seeds, NPO/SimNPO objectives, and Qwen3, where MAST preserves GSM8K while full-parameter unlearning collapses it.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Machine unlearning, one of 12 directions in LLM safety and security. Builds on 16 of its 17 references that are in the map. First posted on arXiv 17 Jun 2026.

Builds on

16 of 17 references.

Closest ancestors · 5

Same direction, newest first

  1. 2025Reviving Your MNEME: Predicting The Side Effects of LLM Unlearning and Fine-Tuning via Sparse Model Diffing19 Jun
  2. 2024Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning9 Oct
  3. 2024Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning8 Apr
  4. 2024Machine Unlearning of Pre-trained Large Language Models23 Feb
  5. 2024TOFU: A Task of Fictitious Unlearning for LLMs11 Jan

Other references · 11

  1. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
  2. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  3. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  4. 2022Mass-Editing Memory in a Transformercited by 1,001
  5. 2025Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?cited by 982
  6. 2024The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearningcited by 509
  7. 2023Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Modelscited by 243
  8. 2025The Path Not Taken: RLVR Provably Learns Off the Principalscited by 56
  9. 2026On the Direction of RLVR Updates for LLM Reasoning: Identification and Exploitationcited by 13
  10. 2026GeoRA: Geometry-Aware Low-Rank Adaptation for RLVRcited by 4
  11. 2026Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimescited by 4

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026REDaCT: Collapse-Free Unlearning in Large Reasoning Models
  2. 2026Selective Transfer of RL Updates for Visual Reasoning
  3. 2026Decouple before Integration: Test-time Synthesis of SFT and RLVR Task Vectors
  4. 2026Detecting RLVR Training Data via Structural Convergence of Reasoning
  5. 2026LEMUR: Latent Entropy-aware Multimodal Unlearning via Visual-anchored Reasoning Redirection
  6. 2026GUARD: Natural Forgetting in Large Reasoning Models via Guided Answer-Reasoning Distillation
  7. 2026Selective Forgetting for Large Reasoning Models
  8. 2026Smart Picks in the Dark: Towards Efficient RLVR for Reasoning via Tracing Metacognitive Pivots