Research map / Reinforcement learning / RL from feedback
Why DPO is a Misspecified Estimator and How to Fix It
Aditya Gopalan, Sayak Ray Chowdhury, Debangshu Banerjee
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 |
|---|---|---|---|
| add auxiliary nullspace variables to the DPO loss (AuxDPO) | DPO, Llama3.1-8B · MMLU-Pro (in-domain) | % accuracy change vs base: 57.14 -> 63.26 | Improved |
| add auxiliary nullspace variables to the DPO loss (AuxDPO) | DPO, Llama3.1-8B · RewardBench v2 (OOD) | % accuracy change vs base: 14.31 -> 32.44 | Improved |
| add auxiliary nullspace variables to the DPO loss (AuxDPO) | DPO, Qwen3-0.6B · RewardBench v2 (OOD) | % accuracy change vs base: -8.16 -> 18.36 | Improved |
| add auxiliary nullspace variables to the DPO loss (AuxDPO) | DPO, Llama3.2-1B · RewardBench v2 (OOD) | % accuracy change vs base: 14.11 -> 43.27 | Improved |
| use AuxDPO instead of IPO | Llama3.2-1B · MMLU-Pro (OOD) | % accuracy change vs base: 14.58 (IPO) vs 12.52 (AuxDPO) | Hurt |
| add auxiliary nullspace variables to the DPO loss (AuxDPO) | DPO, Llama3.1-8B · MMLU-Pro overall accuracy (OOD) | accuracy (%): 27.06 -> 39.26 | Improved |
| add auxiliary nullspace variables to the DPO loss (AuxDPO) | DPO, Llama3.1-8B · MMLU-Pro overall accuracy (in-domain) | accuracy (%): 46.60 -> 51.95 | Improved |
| reduce trainable parameters (LoRA r=4 or last layer only) (ablation) | AuxDPO vs DPO · MMLU-Pro and RewardBench v2 (ID) | accuracy | Improved |
| increase trainable fraction above 75% (ablation) | DPO and AuxDPO · MMLU-Pro and RewardBench v2 | accuracy | No clear effect |
Closest papers by meaning
- As Simple as Fine-tuning: LLM Alignment via Bidirectional Negative Feedback Loss (ICLR 2025 · similarity 0.91)
- Reward Model Learning vs. Direct Policy Optimization: A Comparative Analysis of Learning from Human Preferences (ICML 2024 · similarity 0.91)
- AlphaPO: Reward Shape Matters for LLM Alignment (ICML 2025 · similarity 0.90)
- SimPO: Simple Preference Optimization with a Reference-Free Reward (NeurIPS 2024 · similarity 0.90)
- Is DPO Superior to PPO for LLM Alignment? A Comprehensive Study (ICML 2024 · similarity 0.89)
- Earlier Tokens Contribute More: Learning Direct Preference Optimization From Temporal Decay Perspective (ICLR 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.
