Research map / Large language models / Alignment and preferences
Multi-Objective Preference Optimization: Improving Human Alignment of Generative Models
Akhil Agnihotri, Rahul Jain, Deepak Ramachandran, Zheng Wen
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use the Lagrangian formulation (MOPO-Lag) instead of the log-barrier formulation (MOPO-LB) (ablation) | MOPO · Helpful Assistant helpful-harmless, Zephyr-7b-beta | normalized reward r1: 0.51 -> 0.53 | Improved |
| constrain secondary objectives instead of scalarizing with MODPO | MODPO · Three-objective Helpful Assistant, Zephyr-7b-beta | normalized reward (helpful): 0.04 -> 0.39 | Improved |
| use multi-objective MOPO instead of single-objective RLHF | single-reward RLHF · Three-objective Helpful Assistant, Zephyr-7b-beta | harmless reward: -0.23 -> 0.17 | Improved |
| train DPO only on pairs where both objectives agree (DPO on DJ) versus MOPO | DPO on DJ · Helpful Assistant helpful-harmless, Zephyr-7b-beta | normalized reward r1: 0.53 -> 0.34 | Hurt |
| vary constraint relaxation parameter beta from 0.9 to 0.999 (ablation) | MOPO · Llama-3.1-8B, Helpful Assistant and Reddit Summary | average reward | No clear effect |
| vary lagged-reference interval t0 (1, 100, 1000, never) (ablation) | MOPO · Llama-3.1-8B, Helpful Assistant and Reddit Summary | average reward | No clear effect |
| add MOPO alignment to open-source 7B chat models | Zephyr-7b-beta and other SFT models · Helpful Assistant helpfulness vs harmlessness | helpfulness and harmlessness | Improved |
| impose length constraints as secondary objectives | MOPO · Verbosity-constrained task, Zephyr-7b-beta | response length within range; helpfulness | Improved |
| lagged reference policy updates and adaptive constraint schedule | MOPO · general optimization | optimization stability | Improved |
| MOPO versus PARM on Mistral-7b-v0.2 Instruct helpful-harmless | PARM · Helpful Assistant helpful-harmless, Mistral-7b-v0.2 | normalized reward r1: PARM 0.48±0.03 vs MOPO-Lag 0.45±0.04 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PARM — PARM: Multi-Objective Test-Time Alignment via Preference-Aware Autoregressive Reward Model (ICML 2025) · better on 6 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Robust Multi-Objective Preference Alignment with Online DPO (AAAI 2025 · similarity 0.93)
- Panacea: Pareto Alignment via Preference Adaptation for LLMs (NeurIPS 2024 · similarity 0.91)
- Sequential Preference Optimization: Multi-Dimensional Preference Alignment with Implicit Reward Modeling (AAAI 2025 · similarity 0.91)
- MPO: An Efficient Post-Processing Framework for Mixing Diverse Preference Alignment (ICML 2025 · similarity 0.90)
- Direct Alignment with Heterogeneous Preferences (NeurIPS 2025 · similarity 0.90)
- Linear Alignment: A Closed-form Solution for Aligning Human Preferences without Tuning and Feedback (ICML 2024 · similarity 0.90)
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.
