Research map / Large language models / Alignment and preferences
Beyond Pairwise: Empowering LLM Alignment With (Ranked) Choice Modeling
Yuxuan Tang, Yifan Feng
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 |
|---|---|---|---|
| train on top-2 ranked feedback instead of single-best (ablation) | Mallows-RMJ-PO · AlpacaEval 2 | win rate: 51.17 -> 53.01 | Improved |
| use Mallows-RMJ instead of MNL on pairwise data (ablation) | RCPO · AlpacaEval 2 | win rate: 48.08 -> 48.71 (MNL-PO-Discrete vs Mallows-RMJ-PO-Pairwise) | Improved |
| use Mallows-RMJ instead of MNL on top-2 feedback (ablation) | RCPO · AlpacaEval 2 | win rate: 47.69 -> 53.01 | Improved |
| raise ranked choice length k beyond 2 (assortment size 5) (ablation) | Mallows-RMJ-PO · AlpacaEval 2 | LC win rate: 41.95 -> 39.92 (k=2 to k=4) | Hurt |
| increase assortment size from 2 to 3 (ablation) | Mallows-RMJ-PO · AlpacaEval 2 | win rate: 48.71 -> 50.49 | Improved |
| increase assortment size from 3 to 5 (ablation) | Mallows-RMJ-PO-Top-2 · AlpacaEval 2 | win rate: 49.65 -> 53.01 | Improved |
| increase assortment size from 2 to 5 (ablation) | Mallows-RMJ-PO · Training time | hours: 1.5 -> 3.5 | Hurt |
| increase k (ablation) | Mallows-RMJ-PO · Training time | hours | No clear effect |
| train on ranked choice data instead of converting top-2 ranking to all pairs for DPO | DPO · AlpacaEval 2 | LC win rate: 33.02 -> 41.95 (DPO-AllPairs vs Mallows-RMJ-PO-Top-2) | Improved |
| apply Mallows-RMJ-PO-Top-2 to Gemma-2-9B-it | Gemma-2-9B-it · AlpacaEval 2 | LC win rate: 58.01 (DPO) vs 55.64 | Mixed |
| apply Mallows-RMJ-PO-Top-2 to Mistral-7B-Instruct | Mistral-7B-Instruct · Arena-Hard | win rate: 16.8 (DPO) vs 16.9 | No clear effect |
| Mallows-RMJ-PO-Top-2 vs SimPO on AlpacaEval LC | Llama-3-8B-Instruct · AlpacaEval 2 | LC win rate: 44.15 -> 41.95 | 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 7 of 7 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
- Less is More: Improving LLM Alignment via Preference Data Selection (NeurIPS 2025 · similarity 0.92)
- ComPO: Preference Alignment via Comparison Oracles (NeurIPS 2025 · similarity 0.92)
- Proximalized Preference Optimization for Diverse Feedback Types: A Decomposed Perspective on DPO (NeurIPS 2025 · similarity 0.92)
- Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback (ICML 2025 · similarity 0.92)
- TODO: Enhancing LLM Alignment with Ternary Preferences (ICLR 2025 · similarity 0.92)
- Robust LLM Alignment via Distributionally Robust 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.
