AI Paper Atlas

Research map / Large language models / Alignment and preferences

DRIFT: Learning from Abundant User Dissatisfaction in Real-World Preference Learning

Yifan Wang, Bolian Li, Junlin Wu, Zhaoxuan Tan, Zheli Liu, Ruqi Zhang, Ananth Grama, Qingkai Zeng

ICLR 2026 · Large language models › Alignment and preferences · approach: Reward model design and preference models · 5 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 DRIFT instead of IterDPOQwen2.5-7B-Instruct · WildFeedback controlled iter1 WildBenchTask Score: 47.07 -> 51.06Improved
use DRIFT instead of SPINQwen2.5-7B-Instruct · WildFeedback controlled iter1 WildBenchTask Score: 42.86 -> 51.06Improved
run a second SPIN iterationSPIN · WildFeedback controlled 7B WildBenchTask Score: 42.86 -> 37.86Hurt
train DRIFT on DSAT negatives with fresh policy positives vs baseQwen2.5-14B-Instruct · UltraFeedback WildBenchTask Score: 55.08 -> 59.27Improved
use DRIFT instead of baseQwen2.5-7B-Instruct · AlpacaEval2 WildFeedback controlledwin rate: 37.69 -> 43.79Improved
use full 11k DSAT data instead of controlled 4k for IterDPOIterDPO iter2 · WildFeedback 7B WildBenchTask Score: 48.94 -> 46.17Hurt
use full 11k DSAT data instead of controlled 4k for DRIFTDRIFT iter2 · WildFeedback 7B WildBenchTask Score: 51.06 -> 51.69Improved
remove guidance prompt (DSAT reference and improvement instruction) from DRIFT positive generation (ablation)DRIFT iter1 · WildFeedback full 7B WildBenchTask Score: 50.61 -> 50.77No clear effect
add guidance prompt to IterDPO (ablation)IterDPO iter1 · WildFeedback full 7B WildBenchTask Score: 49.44 -> 46.31Hurt
extend DRIFT from iter4 to iter5DRIFT · WildFeedback 7B five iterationsWildBench Task Score: 52.47 -> 51.22No clear effect
use DRIFT instead of IterDPO on Gemma-3-12B-itGemma-3-12B-it · WildFeedback WildBenchTask Score: 48.98 -> 60.88Improved
train with DRIFT instead of baseQwen2.5-7B-Instruct · ToxiGentoxic score: 1.56 -> 1.58No clear effect

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