AI Paper Atlas

Research map / Large language models / Alignment and preferences

ActiveDPO: Active Direct Preference Optimization for Sample-Efficient Alignment

Xiaoqiang Lin, Arun Verma, Zhongxiang Dai, Daniela Rus, See-Kiong Ng, Bryan Kian Hsiang Low

ICLR 2026 · Large language models › Alignment and preferences · approach: DPO design choices

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
Select preference data with gradient-based uncertainty (ActiveDPO) instead of random/APO/APLPDPO (Llama-2-7B, Gemma-2B, Qwen3-4B) · TLDR and WebGPTaverage rewardImproved
Use heuristic APLP selection on Llama-2DPO · TLDR and WebGPT, Llama-2-7Baverage rewardHurt
Use APLP selection on GemmaDPO · Gemma-2Baverage rewardImproved
Use APO (linear reward assumption)DPO · TLDR and WebGPTaverage rewardMixed
Lower random projection dimension below 8192 (ablation)ActiveDPO · alignment rewardaverage rewardHurt
Normalize LoRA gradients to unit norm (ablation)ActiveDPO · WebGPT, Gemma-2Baverage rewardImproved
Normalize LoRA gradients to unit norm (ablation)ActiveDPO · TLDRaverage rewardNo clear effect
Choose DPO data per model (ablation)Gemma-2B models 1 and 2 · three DPO subsetswin-rateMixed

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