Our new X account is live! Follow @wizwand_team for updates
WorkDL logo mark

Mitigating the Safety Alignment Tax with Null-Space Constrained Policy Optimization

About

As Large Language Models (LLMs) are increasingly deployed in real-world applications, it is important to ensure their behaviors align with human values, societal norms, and ethical principles. However, safety alignment under Reinforcement Learning (RL) often suffers from forgetting learned general abilities, which is also known as the alignment tax. To address this issue, we introduce Null-Space constrained Policy Optimization (NSPO), a novel RL framework for LLM safety alignment while preserving their core abilities. The safety policy gradients are geometrically projected into the null space of general tasks, thereby mitigating the safety alignment tax. In addition, we theoretically prove that NSPO preserves the model's original core capabilities, while still guaranteeing a descent direction for effective safety alignment. Extensive experiments demonstrate that NSPO outperforms existing methods by a large margin, achieving state-of-the-art safety performance without sacrificing accuracy on general tasks, including math, code, and instruction-following tasks. Notably, NSPO is data-efficient and only requires 40% of public human-annotated safety data from PKU-SafeRLHF to achieve promising safety performance, without a large amount of mixed general tasks data in existing alignment methods.

Yifan Niu, Han Xiao, Dongyi Liu, Nuo Chen, Jia Li• 2025

Related benchmarks

TaskDatasetResultRank
General KnowledgeMMLU
MMLU General Knowledge Accuracy71.74
170
Instruction FollowingAlpacaEval
Win Rate94.33
125
Math ReasoningMATH
Accuracy75.05
88
Code GenerationLiveCodeBench
Pass@10.2405
86
Math ReasoningOlympiadBench
Accuracy39
54
Harmful Request DefenseAdvBench
ASR0.06
44
Prohibited Content DetectionALERT
ASR0.01
34
Harmful QueryPKU-Safe
ASR0.52
20
Harmful QueryHarmB
ASR (%)0.18
20
Harmful QueryJailbreakB
ASR0.67
20
Showing 10 of 13 rows

Other info

Follow for update