AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

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

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

ICLR 2026 · Reinforcement learning › RL from feedback · approach: Reward modeling versus direct optimization · 15 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
apply NSPO safety alignment (null-space projected RL)Llama3-8B-Instruct · HarmBenchASR (%): 7.50 -> 0.18Improved
apply NSPO safety alignmentLlama3-8B-Instruct · SORRY-BenchASR (%): 34.15 -> 16.81Improved
apply NSPO safety alignmentQwen2.5-7B-Instruct · PKU-SafeRLHFASR (%): 2.95 -> 0.52Improved
apply NSPO safety alignmentLlama3-8B-Instruct · MMLU / GSM8K / MATHaccuracy (%): GSM8K 75.44 -> 76.42; MATH 27.55 -> 27.10; MMLU 63.79 -> 63.64No clear effect
apply NSPO safety alignmentQwen2.5-7B-Instruct · MMLU / GSM8K / MATHaccuracy (%): GSM8K 81.65 -> 81.96; MATH 74.80 -> 75.05No clear effect
apply NSPO safety alignmentLlama3-8B-Instruct · AlpacaEvalwin rate (%): 96.15 -> 93.23Hurt
use NSPO instead of W-DOORQwen2.5-7B-Instruct · GSM8Kaccuracy (%): 57.85 -> 81.96Improved
add null-space projection and remove KL penalty (NSPO) versus GRPO with KL (ablation)GRPO · safety and general benchmarksASR and accuracyImproved
remove KL penalty from GRPO without projection (ablation)GRPO w/o KL · general capability benchmarksaccuracyHurt
use mixed-domain data for projection matrix instead of single domain (ablation)NSPO · safety and utility averagesutilityImproved
increase projection-matrix sample size from 500 to 2000 (ablation)NSPO · safety and utility averagessafety / utilityMixed
train on 100% instead of 40% of PKU-SafeRLHF (ablation)NSPO · safety performanceperformanceNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy