Mechanism-Guided Selective Unlearning for RLVR-Induced Reasoning
Researchers propose MAST, a mechanism-guided method for selectively unlearning RLVR-induced reasoning in LLMs with less collateral damage than full-parameter updates. Evaluations on Qwen2.5-Math-1.5B and Qwen3-1.7B-Base show MAST preserves performance on MATH and GSM8K benchmarks while reducing damage. This approach helps builders mitigate negative effects of RLVR on reasoning tasks.
Key takeaways
- MAST reduces collateral damage from unlearning RLVR-induced reasoning.
- Preserves performance on MATH and GSM8K benchmarks.
- Evaluated on Qwen2.5-Math-1.5B and Qwen3-1.7B-Base models.
Researchers propose MAST, a mechanism-guided method for selectively unlearning RLVR-induced reasoning in LLMs with less collateral damage than full-parameter updates. Evaluations on Qwen2.5-Math-1.5B and Qwen3-1.7B-Base show MAST preserves performance on MATH and GSM8K benchmarks while reducing damage. This approach helps builders mitigate negative effects of RLVR on reasoning tasks.
Key takeaways
- MAST reduces collateral damage from unlearning RLVR-induced reasoning.
- Preserves performance on MATH and GSM8K benchmarks.
- Evaluated on Qwen2.5-Math-1.5B and Qwen3-1.7B-Base models.