Akari Asai

Akari Asai has made openness a technical proposition rather than a slogan. Her research combines retrieval, reinforcement learning and transparent evaluation to help language models investigate scientific literature, conduct long-form research and know when the available evidence does not support an answer.

Akari Asai’s work is organised around a practical question: can advanced research systems be both useful and inspectable? Trained at the University of Tokyo and the University of Washington, she developed her recent programme across Washington and the Allen Institute for AI, and is due to join Carnegie Mellon University as an assistant professor in autumn 2026. She is first author of OpenScholar, published in Nature in February 2026, a retrieval-augmented system that searches a corpus of 45 million scientific papers before composing cited answers. On the project’s ScholarQABench evaluation, the open eight-billion-parameter model exceeded GPT-4o by 6.1 percentage points and PaperQA2 by 5.5 points for correctness; these are study-specific comparisons, not general capability claims. Asai also co-authored DR Tulu, which introduced reinforcement learning with evolving rubrics for open-ended deep research, and Binary Retrieval-Augmented Reward, which reported lower hallucination rates while preserving performance on instruction following, mathematics and code. The common thread is accountability: release the models, data and evaluation materials, then make evidence—not fluency—the standard against which an answer is judged.