Research map / Large language models / Alignment and preferences
Robust Multi-Objective Preference Alignment with Online DPO
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MO-ODPO, a single prompt-conditioned policy trained with online DPO on sampled objective-weight prefixes, to steer LLM outputs across multiple reward objectives at inference time.
Problem
Multi-objective preference alignment of LLMs needs steerable policies that adapt to variable objective weights at inference, without high training or inference cost, and without multi-objective preference labels.
Approach
Combine prompt-based objective-weight conditioning with online DPO: sample Dirichlet weights per prompt, sample two on-policy responses, score them with reward models, and build preference pairs using the weighted scalar score, training one policy for all weight combinations.
Key results
- MO-ODPO Pareto-dominates all baselines except ODPO Soups on Anthropic-HH and TL;DR with PaLM 2 XS and XXS policies.
- Autorater win rates for MO-ODPO over baselines are 3-15% higher than baselines' win rates, per the paper's reported comparison.
- MO-ODPO takes about 5.2 TPU hours per epoch versus 4.4 for RLFT, while being cheaper than training separate specialist policies.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MO-ODPO instead of the baselines | PaLM 2 XS and XXS policies · Anthropic-HH | Pareto reward tradeoff | Helped |
| Use MO-ODPO instead of the baselines | PaLM 2 XS and XXS policies · TL;DR Summarization | Pareto reward tradeoff | Helped |
| Use MO-ODPO instead of individual baselines | PaLM 2 XS policies · Anthropic-HH and TL;DR Summarization | LLM-judged win rate: 3-15% higher | Helped |
| Increase Dirichlet alpha (ablation) | MO-ODPO · Anthropic-HH | steerability | Helped |
| Decrease Dirichlet alpha (ablation) | MO-ODPO · Anthropic-HH | steerability | Hurt |
| Increase Dirichlet alpha (ablation) | MO-ODPO · TL;DR Summarization | steerability | Hurt |
| Train for additional epochs (ablation) | MO-ODPO · TL;DR Summarization | steerability | Helped |
| Train for additional epochs (ablation) | Rewards-in-Context · TL;DR Summarization | Pareto frontier | No clear effect |
| Apply a response-length penalty (ablation) | MO-ODPO · Anthropic-HH and TL;DR Summarization | Pareto reward tradeoff | Helped |
| Use KL-adjusted rewards (ablation) | MO-ODPO · Anthropic-HH and TL;DR Summarization | Pareto reward tradeoff | Helped |
Surprising findings
- Lower Dirichlet alpha causes bimodal collapse on Anthropic-HH, while higher alpha reduces steerability on TL;DR, so sampling distribution choice matters.
Limitations
- Conditioning on objective weights guides but does not guarantee adherence to the desired objectives.
- Evaluated only with two objectives per benchmark and proprietary PaLM 2 models.
- Gains may partly reflect longer responses, since reward correlates with length.
Experimental setup
Datasets: Anthropic-HH Reddit TL;DR
Benchmarks: Anthropic-HH TL;DR Summarization
Metrics: Pareto front reward trade-off win rate steerability
Baselines: Rewarded Soups ODPO Soups P-MORL Rewards-in-Context MODPO MO-RLHF
Builds on: DPO Rewarded Soups P-MORL Rewards-in-Context MODPO
Closest papers by meaning
- Robust LLM Alignment via Distributionally Robust Direct Preference Optimization (NeurIPS 2025 · similarity 0.92)
- Panacea: Pareto Alignment via Preference Adaptation for LLMs (NeurIPS 2024 · similarity 0.91)
- Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment (NeurIPS 2024 · similarity 0.91)
- MATO: Multi-objective Personalized Alignment with Test-time Optimization for Large Language Models (NeurIPS 2026 · similarity 0.90)
- 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward (ICLR 2025 · similarity 0.90)
- Multi-Reference Preference Optimization for Large Language Models (AAAI 2025 · 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.