Research map / Large language models / Alignment and preferences
Autoregressive Direct Preference Optimization
Masanari Oi, Mahiro Ukai, Masahiro Kaneko, Naoaki Okazaki, Nakamasa Inoue
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 DPO's summed log-ratio with prefix-wise ADPO loss (m=256) | DPO, Llama-3-8B · GSM8K | accuracy: 64.37 -> 68.08 | Improved |
| replace DPO loss with ADPO (m=256) | DPO, Qwen-3-8B · MATH500 | accuracy: 53.80 -> 55.40 | Improved |
| apply ADPO to cDPO (cADPO) | cDPO, Qwen-3-8B · GSM8K | accuracy: 90.98 -> 91.74 | Improved |
| apply ADPO to cDPO (cADPO) | cDPO, Llama-3-8B · MATH500 | accuracy: 16.80 -> 20.20 | Mixed |
| use ADPO instead of DPO on conversation tasks | DPO, Llama3-8B-Instruct · AlpacaEval 2 (LC) | win rate: 40.3 -> 45.8 | Improved |
| use ADPO instead of SimPO | SimPO, Llama3-8B-Instruct · AlpacaEval 2 (LC) | win rate: 44.7 -> 45.8 | Improved |
| use few segments (adaptive m=2) instead of full-response DPO (ablation) | ADPO, Llama-3-8B · GSM8K | accuracy: 64.37 -> 64.29 | No clear effect |
| increase adaptive segments to m>=128 (ablation) | ADPO · GSM8K and MATH500 | accuracy | Improved |
| finer static window (k<=4) vs DPO (ablation) | ADPO static family · GSM8K and MATH500 | accuracy | Improved |
| token-level static ADPO (k=1) (ablation) | ADPO, DeepSeek-Math-7B · GSM8K | accuracy: 67.78 -> 70.36 | Improved |
| adaptive vs static family (ablation) | ADPO · GSM8K and MATH500 | accuracy | Improved |
| increase m from 256 to 512 (ablation) | ADPO adaptive, Llama-3-8B · GSM8K | accuracy: 68.08 -> 67.32 | 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 2 of 2 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
- 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward (ICLR 2025 · similarity 0.92)
- What Matters in Data for DPO? (NeurIPS 2025 · similarity 0.92)
- Earlier Tokens Contribute More: Learning Direct Preference Optimization From Temporal Decay Perspective (ICLR 2025 · similarity 0.91)
- Multi-Reference Preference Optimization for Large Language Models (AAAI 2025 · similarity 0.91)
- Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift (ICML 2025 · similarity 0.91)
- $\beta$-DPO: Direct Preference Optimization with Dynamic $\beta$ (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.
