AI Paper Atlas

Research map / Large language models / Alignment and preferences

Layer-wise Gradient Disentanglement: Decoupling Semantics and Preferences in Direct Preference Optimization

Mengyang Li, Shuang Liu, Zhong Zhang

ICML 2026 · Large language models › Alignment and preferences · approach: Robust and calibrated DPO

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
add GDO bi-dimensional gradient-guided curriculum to DPODPO · UltraFeedback AlpacaEval 2.0, Llama-3-8Bwin rate: 28.4 -> 32.5Improved
add GDO curriculum to SimPOSimPO · UltraFeedback AlpacaEval 2.0, Llama-3-8Bwin rate: 29.8 -> 33.1Improved
add GDO curriculum to IPOIPO · UltraFeedback Arena-Hard, Llama-3-8Bwin rate: 18.1% -> 20.2% (+2.1)Improved
add GDO curriculum to ORPOORPO · UltraFeedback Arena-Hard, Llama-3-8Bwin rate: 18.3% -> 20.0% (+1.7)Improved
use bi-dimensional curriculum instead of a single combined difficulty score (ablation)GDO-DPO · UltraFeedback Arena-Hard, Llama-3-8Bwin rate: 19.5% -> 21.4%Improved
easy-to-hard curriculum direction instead of reversed (hard first) (ablation)GDO-DPO · UltraFeedback Arena-Hard, Llama-3-8Bwin rate: 16.2% -> 21.4%Improved
use gradient-based state-aware pacing instead of fixed linear schedule (ablation)GDO-DPO · UltraFeedback Arena-Hard, Llama-3-8Bwin rate: 19.8% -> 21.4%Improved
use gradient-based pacing instead of loss-based pacing (ablation)GDO-DPO · UltraFeedback Arena-Hard, Llama-3-8Bwin rate: 19.6% -> 21.4%Improved
add the representation stability monitor Srep (ablation)GDO-DPO · UltraFeedback Arena-Hard, Llama-3-8Bwin rate: 20.1% -> 21.4%Improved
add the discrimination readiness monitor Adisc (ablation)GDO-DPO · UltraFeedback Arena-Hard, Llama-3-8Bwin rate: 20.4% -> 21.4%Improved
present high-preference-uncertainty samples first (ablation)DPO · UltraFeedback Arena-Hard, Llama-3-8Bwin rate: 18.5% -> 16.2%Hurt
set layer boundary Lmid to L/2 instead of 2L/3 (ablation)GDO-DPO · UltraFeedback Arena-Hard, Llama-3-8Bwin rate: 21.4% -> 20.5%No clear effect

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.

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

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