Research map / Large language models / Alignment and preferences
Simultaneous Multi-objective Alignment Across Verifiable and Non-verifiable Rewards
Yiran Jenny Shen, Yu Xia, Jonathan Daniel Chang, Prithviraj Ammanabrolu
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-action-head DPO ensemble instead of single-head DPO | DPO on Qwen2.5-7B-Instruct · MATH engagement | engagement: 0.7160 -> 0.8733 | Improved |
| add multi-action-head DPO ensemble instead of MODPO | MODPO · UltraFeedback helpfulness | helpfulness reward: 0.6175 -> 0.6389 | Improved |
| add multi-action-head DPO ensemble instead of single-head DPO | DPO on Llama-3.1-8B-Instruct · UltraFeedback truthfulness | truthfulness reward: 0.2014 -> 0.2478 | Improved |
| add multi-action-head DPO ensemble instead of MODPO on MATH accuracy | MODPO · MATH accuracy | accuracy: 0.7280 -> 0.7247 | No clear effect |
| add accuracy value-guided decoding | base Qwen2.5-7B-Instruct step-wise · MATH accuracy | accuracy: 0.6853 -> 0.7993 | Improved |
| add accuracy value-guided decoding | base Qwen2.5-7B-Instruct step-wise · MATH engagement | engagement: 0.5133 -> 0.4553 | Hurt |
| add engaging PRM-guided decoding | base Qwen2.5-7B-Instruct step-wise · MATH engagement | engagement: 0.5133 -> 0.7187 | Improved |
| add accuracy value PRM guided decoding to MAH-DPO | MAH-DPO Ensemble · MATH accuracy | accuracy: 0.7247 -> 0.8000 | Improved |
| add engaging PRM guidance to MAH-DPO | MAH-DPO Ensemble · MATH engagement | engagement: 0.8733 -> 0.9060 | Improved |
| add honest PRM guidance to MAH-DPO | MAH-DPO Ensemble · UltraFeedback honesty | honesty reward: 0.3687 -> 0.5196 | Improved |
| segment by paragraph break instead of every 64 tokens (ablation) | PRM-guided decoding · MATH accuracy | accuracy (%): 73.0 -> 79.9 | Improved |
| segment by paragraph break instead of every 128 tokens (ablation) | PRM-guided decoding · MATH accuracy | accuracy (%): 77.2 -> 79.9 | Improved |
Closest papers by meaning
- MATO: Multi-objective Personalized Alignment with Test-time Optimization for Large Language Models (NeurIPS 2026 · similarity 0.93)
- Decoding-Time Language Model Alignment with Multiple Objectives (NeurIPS 2024 · similarity 0.92)
- MetaAligner: Towards Generalizable Multi-Objective Alignment of Language Models (NeurIPS 2024 · similarity 0.91)
- Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time (ICML 2025 · similarity 0.91)
- PARM: Multi-Objective Test-Time Alignment via Preference-Aware Autoregressive Reward Model (ICML 2025 · similarity 0.91)
- Panacea: Pareto Alignment via Preference Adaptation for LLMs (NeurIPS 2024 · similarity 0.91)
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.
