Multi-Objective Preference Optimization: Improving Human Alignment of Generative Models

Akhil Agnihotri, Rahul Jain, Deepak Ramachandran, Zheng Wen
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:961-990, 2026.

Abstract

Post-training LLMs with RLHF and preference optimization methods (e.g., DPO, IPO) has greatly improved alignment, yet these approaches assume a single objective. In reality, humans express multiple, often conflicting objectives, such as helpfulness and harmlessness, with no natural scalarization. We study the multi-objective preference alignment problem, where a policy must balance several objectives simultaneously. We propose Multi-Objective Preference Optimization (MOPO), a constrained KL-regularized framework that maximizes a primary objective while enforcing lower bounds on secondary objectives via tunable safety thresholds. MOPO operates directly on pairwise preferences without point-wise rewards, and admits simple closed-form iterative updates. Empirically, MOPO recovers Pareto-optimal policies on synthetic benchmarks and, when fine-tuned on human-preference data, yields multi-billion parameter models that achieve higher rewards and Pareto-dominate baselines, with stable and robust optimization dynamics.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-agnihotri26a, title = {Multi-Objective Preference Optimization: Improving Human Alignment of Generative Models}, author = {Agnihotri, Akhil and Jain, Rahul and Ramachandran, Deepak and Wen, Zheng}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {961--990}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/agnihotri26a/agnihotri26a.pdf}, url = {https://proceedings.mlr.press/v306/agnihotri26a.html}, abstract = {Post-training LLMs with RLHF and preference optimization methods (e.g., DPO, IPO) has greatly improved alignment, yet these approaches assume a single objective. In reality, humans express multiple, often conflicting objectives, such as helpfulness and harmlessness, with no natural scalarization. We study the multi-objective preference alignment problem, where a policy must balance several objectives simultaneously. We propose Multi-Objective Preference Optimization (MOPO), a constrained KL-regularized framework that maximizes a primary objective while enforcing lower bounds on secondary objectives via tunable safety thresholds. MOPO operates directly on pairwise preferences without point-wise rewards, and admits simple closed-form iterative updates. Empirically, MOPO recovers Pareto-optimal policies on synthetic benchmarks and, when fine-tuned on human-preference data, yields multi-billion parameter models that achieve higher rewards and Pareto-dominate baselines, with stable and robust optimization dynamics.} }
Endnote
%0 Conference Paper %T Multi-Objective Preference Optimization: Improving Human Alignment of Generative Models %A Akhil Agnihotri %A Rahul Jain %A Deepak Ramachandran %A Zheng Wen %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-agnihotri26a %I PMLR %P 961--990 %U https://proceedings.mlr.press/v306/agnihotri26a.html %V 306 %X Post-training LLMs with RLHF and preference optimization methods (e.g., DPO, IPO) has greatly improved alignment, yet these approaches assume a single objective. In reality, humans express multiple, often conflicting objectives, such as helpfulness and harmlessness, with no natural scalarization. We study the multi-objective preference alignment problem, where a policy must balance several objectives simultaneously. We propose Multi-Objective Preference Optimization (MOPO), a constrained KL-regularized framework that maximizes a primary objective while enforcing lower bounds on secondary objectives via tunable safety thresholds. MOPO operates directly on pairwise preferences without point-wise rewards, and admits simple closed-form iterative updates. Empirically, MOPO recovers Pareto-optimal policies on synthetic benchmarks and, when fine-tuned on human-preference data, yields multi-billion parameter models that achieve higher rewards and Pareto-dominate baselines, with stable and robust optimization dynamics.
APA
Agnihotri, A., Jain, R., Ramachandran, D. & Wen, Z.. (2026). Multi-Objective Preference Optimization: Improving Human Alignment of Generative Models. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:961-990 Available from https://proceedings.mlr.press/v306/agnihotri26a.html.

Related Material