AI Paper Atlas

Research map / Large language models / Alignment and preferences

Convex Optimization for Alignment and Preference Learning on a Single GPU

Miria Feng, Mert Pilanci

ICML 2026 · Large language models › Alignment and preferences · approach: DPO design choices · 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 COALA instead of ORPOLlama-3.1-8B preference alignment · AlpacaEval2 LC WR, EduFeedbackLC WR (%): 23.87 -> 40.90Improved
use COALA instead of DPOLlama-3.1-8B preference alignment · AlpacaEval2 LC WR, EduFeedbackLC WR (%): 40.68 -> 40.90No clear effect
use COALA instead of DPODolphin-2.6-7B preference alignment · AlpacaEval2 LC WR, EduFeedbackLC WR (%): 34.73 -> 40.81Improved
use COALA instead of DPOLlama-3.1-8B preference alignment · AlpacaEval2 LC WR, IMDbLC WR (%): 21.79 -> 27.64Improved
use COALA instead of DPOMistral-7B preference alignment · AlpacaEval2 LC WR, EduFeedbackLC WR (%): 24.19 -> 24.61No clear effect
use COALA instead of DPO and ORPOhuman preference study · EduFeedback and IMDb, 107 participantshuman win rate (%): EDU 39.1 vs DPO 28.8; IMDb 42.7 vs DPO 24.8Improved
use COALA convex head on frozen features instead of DPO full fine-tuningEduFeedback training · Llama-3.1-8B computeTFLOPs: 10253.37 -> 1805.39Improved
remove reference model (reference-free COALA objective) (ablation)COALA vs DPO · single GPU memoryVRAMImproved
add SFT initialization before preference fine-tuning (ablation)DPO · AlpacaEval2win rateImproved
use COALA instead of DPOreward margin during training · Llama-8B, Mistral-7B, Dolphin-2.6-7Breward margin stabilityImproved
freeze base model and train only a convex headCOALA · nuanced or structurally complex alignmentexpressivenessHurt

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