Research map / Large language models / Alignment and preferences
Multi-objective Large Language Model Alignment with Hierarchical Experts
Zhuo Li, Guodong DU, Weiyang Guo, Yigeng Zhou, Xiucheng Li, Wenya Wang, Fangming Liu, Yequan Wang, Deheng Ye, Min Zhang, Jing Li
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 |
|---|---|---|---|
| use HoE instead of MOD | Llama3.1-8B · Psoups 3-objective, Helpful | reward score: 0.94 -> 3.05 | Improved |
| use HoE instead of PAD | Llama3.1-8B · Psoups 3-objective, HHH | reward score: 1.06 vs 1.03 | No clear effect |
| add one router expert alone to two single-objective LoRA experts (ablation) | HoE · Pareto frontier, two objectives | Pareto frontier coverage | Improved |
| add one LoRA expert to the two fixed experts (ablation) | HoE · Pareto frontier, two objectives | Pareto frontier coverage | Mixed |
| add one LoRA expert and one router expert (ablation) | HoE · Pareto frontier, two objectives | Pareto frontier coverage | Improved |
| add a second LoRA expert (2 LoRA) (ablation) | HoE · Pareto frontier, two objectives | Pareto frontier coverage | Improved |
| lower LoRA rank below 128 (ablation) | HoE · Math task | performance | Hurt |
| use Tchebycheff instead of linear scalarization (ablation) | router training · Pareto frontier stability | training stability | Improved |
| use HoE instead of MOD at inference | Llama2-7B · 3-objective alignment | relative inference cost: 3.10 -> 1.23 | Improved |
| use HoE multi-objective LoRA expert instead of PCB/FR/TIES merging | Chinese/Math/Code LLaMA · GSM8K | accuracy: 56.2 (TIES) vs 50.4 | Hurt |
| rescale task-SVD LoRA with factor 1.9 (ablation) | LLaMA2-7B-Chat Math LoRA · GSM8K | accuracy: 26% -> 68% | Improved |
Closest papers by meaning
- MetaAligner: Towards Generalizable Multi-Objective Alignment of Language Models (NeurIPS 2024 · similarity 0.92)
- MATO: Multi-objective Personalized Alignment with Test-time Optimization for Large Language Models (NeurIPS 2026 · similarity 0.91)
- Aligner: Efficient Alignment by Learning to Correct (NeurIPS 2024 · similarity 0.90)
- Improving Model Alignment Through Collective Intelligence of Open-Source Models (ICML 2025 · similarity 0.90)
- Decoding-Time Language Model Alignment with Multiple Objectives (NeurIPS 2024 · similarity 0.90)
- Alignment of Large Language Models with Constrained Learning (NeurIPS 2025 · similarity 0.89)
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.
