[2602.06869] Uncovering Cross-Objective Interference in Multi-Objective Alignment

Computer Science > Computation and Language arXiv:2602.06869 (cs) [Submitted on 6 Feb 2026 (v1), last revised 6 May 2026 (this version, v2)] Title:Uncovering Cross-Objective Interference in Multi-Objective Alignment Authors:Yining Lu, Meng Jiang View a PDF of the paper titled Uncovering Cross-Objective Interference in Multi-Objective Alignment, by Yining Lu and 1 other authors View PDF HTML (experimental) Abstract:We study a persistent failure mode in multi-objective alignment for large language models (LLMs): training improves performance on only a subset of objectives while causing others to degrade. We formalize this phenomenon as cross-objective interference and conduct the first systematic study across scalarization algorithms, showing that interference is pervasive and exhibits strong model dependence. To explain this phenomenon, we derive a local covariance law showing that an objective improves when its reward exhibits positive covariance with the scalarized score. We extend this analysis to clipped surrogate objectives used in modern alignment, demonstrating that the covariance law remains valid under mild conditions despite clipping. Building on this analysis, we propose Covariance Targeted Weight Adaptation (CTWA), a plug-and-play method that maintains positive covariance between objective rewards and the training signal to effectively mitigate cross-objective interference. Finally, we complement these local improvement conditions with a global convergence analy...