Research map / Large language models / Alignment and preferences
Pre-DPO: Improving Data Utilization in Direct Preference Optimization Using a Guiding Reference Model
Junshu Pan, Wei Shen, Shulin Huang, Qiji Zhou, Yue 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 |
|---|---|---|---|
| re-optimize DPO with a guiding reference model from a first DPO round (Pre-DPO) | DPO · AlpacaEval 2 (Llama3.2-3B-Base) | LC win rate (%): 10.5 -> 12.5 | Improved |
| use SimPO-optimized model as guiding reference for DPO | SimPO · AlpacaEval 2 (Llama3.2-3B-Base) | LC win rate (%): 13.1 -> 18.1 | Improved |
| add guiding reference model from DPO | DPO on Qwen2.5-7B-Instruct · Arena-Hard v0.1 | win rate (%): 62.9 -> 68.8 | Improved |
| add guiding reference model from DPO | DPO on Qwen2.5-7B-Base · Arena-Hard v0.1 | win rate (%): 33.1 -> 32.6 | Hurt |
| train DPO for 2 epochs to match Pre-DPO compute (ablation) | DPO on Llama3.2-3B-Base · AlpacaEval 2 | LC win rate (%): 10.5 -> 11.0 | No clear effect |
| train DPO for 2 epochs to match Pre-DPO compute (ablation) | DPO on Llama3.2-3B-Instruct · AlpacaEval 2 | LC win rate (%): 36.3 -> 35.2 | Hurt |
| use guiding reference from DPO after 1 epoch instead of 2nd DPO epoch (ablation) | DPO on Llama3.2-3B-Instruct · AlpacaEval 2 | LC win rate (%): 35.2 -> 39.3 | Improved |
| use ORPO-optimized model as guiding reference | ORPO · AlpacaEval 2 (Llama3.2-3B-Base) | LC win rate (%): 10.2 -> 12.3 | Improved |
| use sDPO-optimized model as guiding reference | sDPO · AlpacaEval 2 (Llama3.2-3B-Base) | LC win rate (%): 12.0 -> 12.9 | Improved |
| use TR-DPO-optimized model as guiding reference | TR-DPO · AlpacaEval 2 (Llama3.2-3B-Base) | LC win rate (%): 11.7 -> 12.8 | Improved |
| use SimPO guiding reference | SimPO on Qwen2.5-7B-Base · AlpacaEval 2 (response length) | avg response length: 1836 -> 1758 | Improved |
| hard reference update every 32 steps (TR-DPO) vs guiding reference (ablation) | DPO on Llama3.2-3B-Base · lambda distribution | weight spread | 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 5 of 5 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
- Explicit Preference Optimization: No Need for an Implicit Reward Model (ICML 2025 · similarity 0.93)
- Earlier Tokens Contribute More: Learning Direct Preference Optimization From Temporal Decay Perspective (ICLR 2025 · similarity 0.93)
- SimPO: Simple Preference Optimization with a Reference-Free Reward (NeurIPS 2024 · similarity 0.93)
- Multi-Reference Preference Optimization for Large Language Models (AAAI 2025 · similarity 0.92)
- MallowsPO: Fine-Tune Your LLM with Preference Dispersions (ICLR 2025 · similarity 0.92)
- Active Preference Learning for Large Language Models (ICML 2024 · similarity 0.92)
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.
