Preference Optimization for Review Question Generation Improves Writing Quality
New model IntelliAsk improves review question quality by 3.6% on writing benchmarks and 3.6 points on reasoning tasks.
Deep Dive
Researchers led by Karun Sharma developed IntelliAsk, a question-generation model trained using their novel IntelliReward reward model and DAPO optimization. Built on Qwen3-32B, it generates more substantive, evidence-based peer review questions by learning from expert human preferences. IntelliAsk shows measurable gains, scoring 68.3 vs 64.7 on MuSR reasoning and 8.31 vs 8.07 on WritingBench. The team released their models and annotations to benchmark LLM-generated review questions.
Why It Matters
Automates high-quality academic peer review, potentially accelerating scientific publishing while maintaining rigorous standards.