Research map / Large language models / Alignment and preferences
AMaPO: Adaptive Margin-attached Preference Optimization for Language Model Alignment
Ruibo Deng, Duanyu Feng, Wenqiang Lei
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 |
|---|---|---|---|
| replace SimPO static margin with AMaPO adaptive margin | SimPO · AlpacaEval 2, Llama3-8B-Base | LC win rate (%): 22.0 -> 26.4 | Improved |
| replace SimPO static margin with AMaPO adaptive margin | SimPO · AlpacaEval 2, Mistral-7B-Instruct | LC win rate (%): 32.1 -> 34.5 | Improved |
| replace SimPO static margin with AMaPO adaptive margin | SimPO · MT-Bench, Llama3-8B-Base | GPT-4 Turbo score: 6.6 -> 6.4 | Hurt |
| replace SimPO static margin with AMaPO adaptive margin | SimPO · RM-Bench Hard, Llama3-8B-Base | ranking accuracy (%): 23.3 -> 25.4 | Mixed |
| replace SimPO static margin with AMaPO adaptive margin | SimPO · UltraFeedback Mutual-OOD, Mistral-7B-Base | ranking accuracy (%): 75.94 -> 82.97 | Improved |
| replace SimPO static margin with AMaPO adaptive margin | SimPO · UltraFeedback ID, Llama3-8B-Base | ranking accuracy (%): 78.1 -> 77.33 | No clear effect |
| remove Z-normalization (ablation) | AMaPO · AlpacaEval 2, Llama3-8B-Base | LC win rate (%): 26.3 -> 24.8 when removed | Improved |
| add exponential scaling (ablation) | AMaPO · AlpacaEval 2 WR, Llama3-8B-Base | win rate (%): 21.4 -> 17.6 when removed | Improved |
| add adaptive margin (Z-norm plus exp scaling) (ablation) | AMaPO · AlpacaEval 2, Llama3-8B-Base | LC win rate (%): 26.3 -> 20.7 when removed | Improved |
| add adaptive margin (Z-norm plus exp scaling) (ablation) | AMaPO · AlpacaEval 2, Llama3-8B-Instruct | LC win rate (%): 46.1 -> 47.2 when removed | Hurt |
| add zero margin for correctly ranked samples (ablation) | AMaPO · MT-Bench, Llama3-8B-Instruct | GPT-4 Turbo score: 7.2 -> 6.7 when removed | Improved |
| add zero margin for correctly ranked samples (ablation) | AMaPO · AlpacaEval 2 LC, Llama3-8B-Instruct | LC win rate (%): 46.1 -> 48.6 when removed | Hurt |
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 10 of 10 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
- SimPER: A Minimalist Approach to Preference Alignment without Hyperparameters (ICLR 2025 · similarity 0.91)
- ROPO: Robust Preference Optimization for Large Language Models (ICML 2025 · similarity 0.91)
- ComPO: Preference Alignment via Comparison Oracles (NeurIPS 2025 · similarity 0.91)
- Cal-DPO: Calibrated Direct Preference Optimization for Language Model Alignment (NeurIPS 2024 · similarity 0.90)
- Less is More: Improving LLM Alignment via Preference Data Selection (NeurIPS 2025 · similarity 0.90)
- RePO: Understanding Preference Learning Through ReLU-Based Optimization (NeurIPS 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.
