AI Paper Atlas

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

ICML 2026 · Large language models › Alignment and preferences · approach: Active and token-level preference learning · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
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 UFLC win rate (%): 14.8 -> 24.1Improved
use separate SFT stage then alignment loss only instead of one-stage from base (ablation)ASFT · AlpacaEval 2 LC, Llama 3.1 8B UFLC win rate (%): 14.5 -> 16.4Improved
drop SFT loss term when aligning from SFT init (ablation)ORPO · AlpacaEval 2 LC, Llama 3.1 8B UFLC win rate (%): 13.4 -> 24.1Improved
drop SFT loss term when aligning from SFT init (ablation)ASFT · AlpacaEval 2 LC, Llama 3.1 8B UFLC win rate (%): 11.4 -> 16.4Improved
add beta temperature to ORPO alignment loss (ablation)ORPO · Llama 3.2 3B TL;DRGPT-4 win rate: +7.0Improved
add beta temperature to ASFT alignment loss (ablation)ASFT · Llama 3.2 3B TL;DRGPT-4 win rate: +43.4Improved
tune beta for ORPO (ablation)ORPO · Llama 3.1 8B UF AlpacaEval 2 LCLC win rate (%): +3.46Improved
tune beta for ASFT (ablation)ASFT · Llama 3.1 8B UF AlpacaEval 2 LCLC win rate (%): +8.27Improved
use pairwise instead of pointwise ranking objectiveDAAs (DPO/IPO/SimPO/ORPO vs APO-Zero/NCA/Cal-DPO/ASFT) · Llama 3.1 8B UF AlpacaEval 2 LCLC win rate (%): 23.15 -> 28.25Improved
use pairwise instead of pointwise ranking objectiveDAAs · Llama 3.2 3B UF AlpacaEval 2 LCLC win rate (%): 10.36 vs 11.43No clear effect
use pairwise instead of pointwise ranking objectiveDAAs · Mistral 7B UFAlpacaEval 2 LC: 21.46 -> 24.52Improved
use pairwise AlphaPO score instead of pointwiseAlphaPO · Qwen2.5-7B Math CoTmean avg@K: 0.2922 -> 0.2945Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy