Research map / Large language models / Alignment and preferences
Alignment through Meta-Weighted Online Sampling: Bridging the Gap between Data Generation and Preference Optimization
Junming Yang, Ning Xu, Biao Liu, Shiqi Qiao, Xin Geng
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 meta-weighted adaptive sampling instead of random sampling (ablation) | MetaAPO (Llama-3.1-8B, DPO) · AlpacaEval 2 | LC win rate (%): 38.53 -> 43.21 | Improved |
| add meta-weighted adaptive sampling instead of threshold sampling (ablation) | MetaAPO (Llama-3.1-8B, DPO) · AlpacaEval 2 | LC win rate (%): 42.92 -> 43.21 | Improved |
| add meta-weighted selective sampling instead of sampling all prompts online (ablation) | MetaAPO (Llama-3.1-8B, DPO) · AlpacaEval 2 | LC win rate (%): 42.04 -> 43.21 | Improved |
| add meta-weighted selective sampling instead of sampling all prompts online (ablation) | MetaAPO (Llama-3.1-8B, DPO) · Annotation ratio | online annotation ratio (%): 100.0 -> 58.4 | Improved |
| add learned meta-weights for loss instead of uniform 0.5 weighting (ablation) | MetaAPO (Llama-3.1-8B, DPO) · AlpacaEval 2 | LC win rate (%): 37.70 -> 43.21 | Improved |
| add learned meta-weights for loss instead of uniform 0.5 weighting (ablation) | MetaAPO (Llama-3.1-8B, DPO) · AlpacaEval 2 | raw win rate (%): 39.25 -> 47.48 | Improved |
| add embedding similarity of chosen/rejected as meta-learner input (ablation) | MetaAPO (Llama-3.1-8B, DPO) · AlpacaEval 2 | LC win rate (%): 43.21 -> 42.88 | No clear effect |
| add continual meta-learner updates instead of freezing the learned meta-learner (ablation) | MetaAPO (Llama-3.1-8B, DPO) · AlpacaEval 2 | LC win rate (%): 40.03 -> 43.21 | Improved |
| add trainable meta-learner instead of fixed sigmoid heuristic (ablation) | MetaAPO (Llama-3.1-8B, DPO) · AlpacaEval 2 | raw win rate (%): 43.07 -> 47.48 | Improved |
| add trainable meta-learner instead of fixed sigmoid heuristic (ablation) | MetaAPO (Llama-3.1-8B, DPO) · Arena-Hard | win rate (%): 38.3 -> 43.9 | Improved |
| add MetaAPO on top of Online DPO | Online DPO (Llama-3.1-8B) · AlpacaEval 2 | LC win rate (%): 39.98 -> 43.21 | Improved |
| MetaAPO versus PPO with fewer online samples | PPO (Llama-3.1-8B) · Arena-Hard | win rate (%): 40.8 -> 43.9 | Improved |
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.
- SimPO — SimPO: Simple Preference Optimization with a Reference-Free Reward (NeurIPS 2024) · better on 4 of 4 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
- Meta-Learning Preferences for Multilingual LLM Alignment (NeurIPS 2026 · similarity 0.92)
- Spread Preference Annotation: Direct Preference Judgment for Efficient LLM Alignment (ICLR 2025 · similarity 0.91)
- MetaAligner: Towards Generalizable Multi-Objective Alignment of Language Models (NeurIPS 2024 · similarity 0.91)
- Leveraging robust optimization for llm alignment under distribution shifts (NeurIPS 2025 · similarity 0.91)
- Online Preference Alignment for Language Models via Count-based Exploration (ICLR 2025 · similarity 0.91)
- Adaptive Batch-Wise Sample Scheduling for Direct Preference Optimization (NeurIPS 2025 · 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.
