Research map / Large language models / Alignment and preferences
Convex Optimization for Alignment and Preference Learning on a Single GPU
Miria Feng, Mert Pilanci
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 |
|---|---|---|---|
| use COALA instead of ORPO | Llama-3.1-8B preference alignment · AlpacaEval2 LC WR, EduFeedback | LC WR (%): 23.87 -> 40.90 | Improved |
| use COALA instead of DPO | Llama-3.1-8B preference alignment · AlpacaEval2 LC WR, EduFeedback | LC WR (%): 40.68 -> 40.90 | No clear effect |
| use COALA instead of DPO | Dolphin-2.6-7B preference alignment · AlpacaEval2 LC WR, EduFeedback | LC WR (%): 34.73 -> 40.81 | Improved |
| use COALA instead of DPO | Llama-3.1-8B preference alignment · AlpacaEval2 LC WR, IMDb | LC WR (%): 21.79 -> 27.64 | Improved |
| use COALA instead of DPO | Mistral-7B preference alignment · AlpacaEval2 LC WR, EduFeedback | LC WR (%): 24.19 -> 24.61 | No clear effect |
| use COALA instead of DPO and ORPO | human preference study · EduFeedback and IMDb, 107 participants | human win rate (%): EDU 39.1 vs DPO 28.8; IMDb 42.7 vs DPO 24.8 | Improved |
| use COALA convex head on frozen features instead of DPO full fine-tuning | EduFeedback training · Llama-3.1-8B compute | TFLOPs: 10253.37 -> 1805.39 | Improved |
| remove reference model (reference-free COALA objective) (ablation) | COALA vs DPO · single GPU memory | VRAM | Improved |
| add SFT initialization before preference fine-tuning (ablation) | DPO · AlpacaEval2 | win rate | Improved |
| use COALA instead of DPO | reward margin during training · Llama-8B, Mistral-7B, Dolphin-2.6-7B | reward margin stability | Improved |
| freeze base model and train only a convex head | COALA · nuanced or structurally complex alignment | expressiveness | Hurt |
Closest papers by meaning
- Earlier Tokens Contribute More: Learning Direct Preference Optimization From Temporal Decay Perspective (ICLR 2025 · similarity 0.90)
- Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints (ICLR 2024 · similarity 0.90)
- Towards Efficient Exact Optimization of Language Model Alignment (ICML 2024 · similarity 0.90)
- As Simple as Fine-tuning: LLM Alignment via Bidirectional Negative Feedback Loss (ICLR 2025 · similarity 0.90)
- 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward (ICLR 2025 · similarity 0.90)
- Active Preference Learning for Large Language Models (ICML 2024 · similarity 0.90)
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.
