Research map / Large language models / Alignment and preferences
Transitivity Meets Cyclicity: Explicit Preference Decomposition for Dynamic Large Language Model Alignment
Yucong Huang, Xiucheng Li, Kaiqi Zhao, 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 |
|---|---|---|---|
| replace BT/GPM with HRC (dim 2+1) as preference model | reward model · RewardBench 2 average | accuracy: 56.40 -> 57.63 | Improved |
| replace BT/GPM with HRC (dim 2+1) as preference model | Llama-3.1-8B-Instruct reward model · RewardBench 2 average | accuracy: 70.10 -> 70.95 | Improved |
| use HRC instead of BT or GPM | reward model, Llama-3.1-8B-Instruct · RewardBench 2 Ties | accuracy: 73.53 -> 74.51 | Improved |
| add context-aware gating (ablation) | HRC (Gemma-2B-it, dim 2+1) · RewardBench 2 average | accuracy: 56.49 -> 57.63 | Improved |
| add reward clipping (ablation) | HRC (Gemma-2B-it, dim 4+1) · RewardBench 2 average | accuracy: 57.24 -> 57.12 | Hurt |
| add unit-norm constraint (ablation) | HRC (Llama-3.1-8B-Instruct, dim 4+1) · RewardBench 2 average | accuracy: 71.25 -> 70.67 | Hurt |
| add unit-norm constraint (ablation) | HRC (Llama-3.1-8B-Instruct, dim 2+1) · RewardBench 2 average | accuracy: 69.98 -> 70.95 | Improved |
| add context-aware gating (ablation) | HRC (Llama-3.1-8B-Instruct, dim 4+1) · RewardBench 2 average | accuracy: 71.04 -> 70.67 | Hurt |
| switch from static SPPO to DSPPO with HRC oracle | Gemma-2B-it alignment · AlpacaEval 2.0 | LC win rate (%): 43.00 -> 44.75 | Improved |
| switch from static SPPO to DSPPO with HRC oracle | Llama-3.1-8B-Instruct alignment, iteration 2 · MT-Bench | score: 8.42 -> 8.14 | Hurt |
| use HRC+DSPPO instead of BT+SPPO or GPM+SPPO | Gemma-2B-it alignment · Arena-Hard-v0.1 | win rate (%): 42.1 -> 46.8 | Improved |
| transfer HRC+DSPPO preference models to Gemma-2-9B-it | Gemma-2-9B-it alignment · AlpacaEval 2.0 | LC win rate (%): 48.79 -> 52.20 | Improved |
Closest papers by meaning
- Self-Play Preference Optimization for Language Model Alignment (ICLR 2025 · similarity 0.93)
- AlphaDPO: Adaptive Reward Margin for Direct Preference Optimization (ICML 2025 · similarity 0.90)
- Iterative Nash Policy Optimization: Aligning LLMs with General Preferences via No-Regret Learning (ICLR 2025 · similarity 0.90)
- Magnetic Preference Optimization: Achieving Last-iterate Convergence for Language Model Alignment (ICLR 2025 · similarity 0.90)
- Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback (ICML 2025 · similarity 0.89)
- Sequential Preference Optimization: Multi-Dimensional Preference Alignment with Implicit Reward Modeling (AAAI 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.
