AI Paper Atlas

Research map / Large language models / Alignment and preferences

Towards Understanding Valuable Preference Data for Large Language Model Alignment

Zizhuo Zhang, Qizhou Wang, Shanshan Ye, Jianing Zhu, Jiangchao Yao, Bo Han, Masashi Sugiyama

ICLR 2026 · Large language models › Alignment and preferences · approach: Active and token-level preference learning · 4 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
add LossDiff-IRM selection (about 50-64% of data) instead of full dataDPO · Llama-3.1-8B, AlpacaEvalWinRate: 78.41 -> 87.08Improved
select data with LossDiff-IRM instead of full dataDPO · Qwen3-8B-Base, AlpacaEvalWinRate: 63.85 -> 71.52Improved
train only on the pairs dropped by LossDiff-IRM instead of full data (ablation)DPO · Llama-3.1-8B, AlpacaEvalWinRate: 78.41 -> 61.15Hurt
train only on dropped pairs instead of full data (ablation)SLiC · Qwen3-8B-Base, Vicuna-BenchWinRate: 54.69 -> 56.72Hurt
use LossDiff-IRM selection instead of the best baselineDPO · Llama-3.1-8B, four benchmarksWinRate: +6.45% on averageImproved
use LossDiff-IRM selection instead of the best baselineDPO · Llama-3.1-8B and Qwen3-8B-BaseWinRate: +2.62%Improved
combine LossDiff and IRM instead of LossDiff alone or IRM alone (ablation)LossDiff-IRM selection · DPO and SLiCWinRateImproved
combine LossDiff and IRM instead of either alone (ablation)LossDiff-IRM · Qwen3-0.6B-Base epoch 1, overlap with exact TIFoverlap coefficient: 0.6953 -> 0.7820Improved
replace exact IF with LossDiff-IRM for scoringdata selection · Llama-3.2-1B probing setcompute time: about 10 hours -> about 5 minutesImproved
flip validation labels at increasing noise rate (ablation)LossDiff-IRM · validation noise 0.1 to 0.4performanceHurt
tighten the percentile thresholds to filter more data (ablation)LossDiff-IRM · Llama-3.1-8B DPO, four benchmarksperformanceMixed
train on small-IF data instead of medium-IF data (ablation)DPO · Qwen3-0.6B-Base and Llama-3.2-1B probing seteval loss and reward marginHurt

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