Research map / Large language models / Alignment and preferences
The Differences Between Direct Alignment Algorithms are a Blur
Alexey Gorbatovski, Boris Shaposhnikov, Viacheslav Sinii, Alexey Malakhov, Daniil Gavrilov
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 separate SFT stage then alignment loss only (SFT init) instead of one-stage with SFT loss from base (ablation) | ORPO · AlpacaEval 2 LC, Llama 3.1 8B UF | LC win rate (%): 14.8 -> 24.1 | Improved |
| use separate SFT stage then alignment loss only instead of one-stage from base (ablation) | ASFT · AlpacaEval 2 LC, Llama 3.1 8B UF | LC win rate (%): 14.5 -> 16.4 | Improved |
| drop SFT loss term when aligning from SFT init (ablation) | ORPO · AlpacaEval 2 LC, Llama 3.1 8B UF | LC win rate (%): 13.4 -> 24.1 | Improved |
| drop SFT loss term when aligning from SFT init (ablation) | ASFT · AlpacaEval 2 LC, Llama 3.1 8B UF | LC win rate (%): 11.4 -> 16.4 | Improved |
| add beta temperature to ORPO alignment loss (ablation) | ORPO · Llama 3.2 3B TL;DR | GPT-4 win rate: +7.0 | Improved |
| add beta temperature to ASFT alignment loss (ablation) | ASFT · Llama 3.2 3B TL;DR | GPT-4 win rate: +43.4 | Improved |
| tune beta for ORPO (ablation) | ORPO · Llama 3.1 8B UF AlpacaEval 2 LC | LC win rate (%): +3.46 | Improved |
| tune beta for ASFT (ablation) | ASFT · Llama 3.1 8B UF AlpacaEval 2 LC | LC win rate (%): +8.27 | Improved |
| use pairwise instead of pointwise ranking objective | DAAs (DPO/IPO/SimPO/ORPO vs APO-Zero/NCA/Cal-DPO/ASFT) · Llama 3.1 8B UF AlpacaEval 2 LC | LC win rate (%): 23.15 -> 28.25 | Improved |
| use pairwise instead of pointwise ranking objective | DAAs · Llama 3.2 3B UF AlpacaEval 2 LC | LC win rate (%): 10.36 vs 11.43 | No clear effect |
| use pairwise instead of pointwise ranking objective | DAAs · Mistral 7B UF | AlpacaEval 2 LC: 21.46 -> 24.52 | Improved |
| use pairwise AlphaPO score instead of pointwise | AlphaPO · Qwen2.5-7B Math CoT | mean avg@K: 0.2922 -> 0.2945 | Improved |
Closest papers by meaning
- Style Outweighs Substance: Failure Modes of LLM Judges in Alignment Benchmarking (ICLR 2025 · similarity 0.87)
- Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples (ICML 2025 · similarity 0.87)
- LLM Safety Alignment is Divergence Estimation in Disguise (NeurIPS 2025 · similarity 0.87)
- Aligner: Efficient Alignment by Learning to Correct (NeurIPS 2024 · similarity 0.86)
- ComPO: Preference Alignment via Comparison Oracles (NeurIPS 2025 · similarity 0.86)
- TODO: Enhancing LLM Alignment with Ternary Preferences (ICLR 2025 · similarity 0.85)
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.
