Research map / Large language models / Alignment and preferences
RMO: Towards Better LLM Alignment via Reshaping Reward Margin Distributions
Yanchi Ru, Yue Huang, Xiangliang Zhang
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 |
|---|---|---|---|
| build batches that maximize intra-batch reward margin variance (High-Var) instead of random (ablation) | DPO · AlpacaEval 2.0 on Anthropic HH, Llama-3.2-1B-Instruct | LC win rate (%): 54.00 -> 56.39 | Improved |
| build batches that minimize margin variance (Low-Var) instead of random (ablation) | DPO · AlpacaEval 2.0 on Anthropic HH, Llama-3.2-1B-Instruct | LC win rate (%): 54.00 -> 49.67 | Hurt |
| add Dual Denoising Filtering (6% filter ratio) (ablation) | DPO with random batches · AlpacaEval 2.0 on Anthropic HH, Llama-3.2-1B-Instruct | LC win rate (%): 54.00 -> 56.05 | Improved |
| add Pairwise Margin Amplification (lambda_reg=2) (ablation) | DPO with high-variance batches · AlpacaEval 2.0 on Anthropic HH, Pythia-1.4B | LC win rate (%): 59.20 -> 62.34 | Improved |
| add Dual Denoising Filtering to high-variance DPO (ablation) | DPO with High-Var batches · AlpacaEval 2.0 on Anthropic HH, Llama-3.2-1B-Instruct | LC win rate (%): 56.39 -> 56.43 | No clear effect |
| add full RMO to high-variance DPO | DPO · AlpacaEval 2.0 on Anthropic HH, Llama-3.2-1B-Instruct | LC win rate (%): 56.39 -> 59.78 | Improved |
| add RMO to SimPO | SimPO · AlpacaEval 2.0 on UltraFeedback, Llama-3.2-3B-Instruct | LC win rate (%): 65.91 -> 69.52 | Improved |
| add Pairwise Margin Amplification to RMO on SimPO (ablation) | RMO (SimPO) · AlpacaEval 2.0 on UltraFeedback, Llama-3.2-3B-Instruct | LC win rate (%): 68.32 -> 69.52 | Improved |
| add Batch Margin Diversification to SimPO (ablation) | SimPO · AlpacaEval 2.0 on Anthropic HH, Llama-3.2-3B-Instruct | LC win rate (%): 60.00 -> 61.44 | Mixed |
| increase lambda_reg beyond 2 (ablation) | Pairwise Margin Amplification · AlpacaEval 2.0, Llama-3.2-1B-Instruct | LC and WR | Hurt |
| increase lambda_reg (ablation) | Pairwise Margin Amplification · Preference accuracy, Llama-3.2-1B-Instruct | accuracy | Improved |
| increase filter ratio beyond 10% (ablation) | Dual Denoising Filtering · AlpacaEval 2.0, Llama-3.2-1B-Instruct | LC win rate (%) | Hurt |
Closest papers by meaning
- Larger or Smaller Reward Margins to Select Preferences for LLM Alignment? (ICML 2025 · similarity 0.93)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 2025 · similarity 0.92)
- Less is More: Improving LLM Alignment via Preference Data Selection (NeurIPS 2025 · similarity 0.92)
- RMB: Comprehensively benchmarking reward models in LLM alignment (ICLR 2025 · similarity 0.92)
- Leveraging robust optimization for llm alignment under distribution shifts (NeurIPS 2025 · similarity 0.92)
- Noise Contrastive Alignment of Language Models with Explicit Rewards (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.
