Research map / Large language models / Alignment and preferences
ActiveUltraFeedback: Efficient Preference Data Generation using Active Learning
Davit Melikidze, Marian Schneider, Jessica Lam, Martin Wertich, Ido Hakimi, Barna Pásztor, Andreas Krause
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 changed | Where | Measured | Effect |
|---|---|---|---|
| select pairs with DRTS instead of random sampling (ablation) | ActiveUltraFeedback with DPO · 4 downstream benchmarks mean (UltraFeedback prompts) | mean score delta vs base: +0.046 -> +0.127 | Improved |
| select pairs with DeltaUCB instead of random sampling (ablation) | ActiveUltraFeedback with DPO · RewardBench 2 | score delta vs base: +0.278 -> +0.339 | Improved |
| use dueling bandit selection (DTS, MaxMinLCB, InfoMax) instead of random (ablation) | ActiveUltraFeedback · 4 downstream benchmarks mean (DPO) | mean score delta vs base: +0.046 (Random) -> +0.023 (DTS), +0.016 (MaxMinLCB) | Hurt |
| use DeltaQwen fixed small-vs-large pairing instead of active selection for reward modeling (ablation) | preference data generation · RewardBench 2 | score delta vs base: +0.100 (DeltaQwen) vs +0.278 (Random) | Hurt |
| use DRTS or DeltaUCB selection with a small subset of data (ablation) | DPO fine-tuning · downstream mean, UltraFeedback prompts | mean score delta | Improved |
| switch DPO to IPO or SimPO with DeltaQwen data (ablation) | DeltaQwen · mean downstream score | mean score delta | Hurt |
| use DRTS or DeltaUCB on other prompt sources (Skywork, Combined, Tulu 3) (ablation) | ActiveUltraFeedback · DPO mean score | mean score delta | Improved |
| use DRTS or DeltaUCB instead of original Skywork/Combined dataset for reward modeling (ablation) | ActiveUltraFeedback · RewardBench 2, Skywork and Combined prompts | score delta | Hurt |
| use DRTS or DeltaUCB with IPO and SimPO instead of DPO (ablation) | ActiveUltraFeedback · mean downstream score, IPO and SimPO | mean score delta | Improved |
| use probabilistic judge scoring (expectation over score tokens) instead of discrete generation (ablation) | LLM judge annotation · UltraFeedback judged pairs | win rate: 75.36% -> 76.70% | Improved |
| enable explicit reasoning in the LLM judge (ablation) | LLM judge annotation · UltraFeedback judged pairs | win rate: 76.70% -> 73.54% | Hurt |
| use probabilistic instead of discrete scoring (ablation) | LLM judge annotation · RewardBench 2 | mean: 0.698 -> 0.744 | Improved |
Closest papers by meaning
- Active Preference Learning for Large Language Models (ICML 2024 · similarity 0.89)
- Deep Bayesian Active Learning for Preference Modeling in Large Language Models (NeurIPS 2024 · similarity 0.89)
- HelpSteer3-Preference: Open Human-Annotated Preference Data across Diverse Tasks and Languages (NeurIPS 2025 · similarity 0.89)
- Active Reward Modeling: Adaptive Preference Labeling for Large Language Model Alignment (ICML 2025 · similarity 0.89)
- MAPLE: A Framework for Active Preference Learning Guided by Large Language Models (AAAI 2025 · similarity 0.88)
- Online Preference Alignment for Language Models via Count-based Exploration (ICLR 2025 · similarity 0.88)
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.
