# Hanna Hajishirzi

> Source: https://aiwiki.ai/wiki/hanna_hajishirzi
> Updated: 2026-07-24
> Categories: AI Research, Computer Vision, Natural Language Processing, People
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Hanna Hajishirzi." aiwiki.ai, 24 Jul 2026. https://aiwiki.ai/wiki/hanna_hajishirzi
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

Hanna Hajishirzi, whose scholarly publications generally use the name Hannaneh Hajishirzi, is a computer scientist whose research covers [natural language processing](https://aiwiki.ai/wiki/natural_language_processing), [machine learning](https://aiwiki.ai/wiki/machine_learning), [generative AI](https://aiwiki.ai/wiki/generative_ai), question answering, and multimodal systems. She is a professor in the Paul G. Allen School of Computer Science & Engineering at the University of Washington and directs the H2Lab. In 2026 she joined Microsoft AI's Superintelligence team while retaining her university faculty position.[1][2][4]

Before joining Microsoft, Hajishirzi was a senior research director at the [Allen Institute for AI](https://aiwiki.ai/wiki/allen_institute_for_ai) (Ai2). Her work there included leadership of [OLMo](https://aiwiki.ai/wiki/olmo), a family of language models released with training data and development tools, and Tülu, a project on open post-training methods. Her academic work has moved from probabilistic reasoning and multimodal interaction to machine reading, instruction following, retrieval, model evaluation, and open language-model infrastructure.[1][2][13]

## Education and early research

Hajishirzi earned a bachelor's degree in computer engineering, with a software specialization, from Sharif University of Technology. Her current curriculum vitae does not give a graduation year for that degree. She completed a Ph.D. in computer science at the University of Illinois at Urbana-Champaign in September 2011. Her dissertation, *Action-centered Reasoning for Probabilistic Dynamic Systems*, was supervised by Eyal Amir; the doctoral committee also included Dan Roth, Julia Hockenmaier, and Erik T. Mueller.[2][3][6]

The dissertation belongs to Hajishirzi's early work on [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) methods for reasoning under uncertainty. Papers from this period studied stochastic filtering, sensing decisions, probabilistic action models, and the interpretation of narratives about dynamic events. The research combined symbolic representations of actions with probabilistic inference, an approach she later applied to language and multimodal problems.[2][6]

From September 2011 to September 2012, Hajishirzi was a postdoctoral associate at Disney Research in Pittsburgh. University of Washington biographies also describe this year as postdoctoral work associated with Carnegie Mellon University. One project from that period examined how an interactive character could determine whether a child in a group was addressing the character. The resulting SIGDIAL 2012 paper modeled audio, visual, conversational, and temporal information and received the conference's best paper award.[2][3][7]

## Academic and research appointments

Hajishirzi joined the University of Washington's electrical engineering department as a research scientist in September 2012. She became a research assistant professor in January 2015, then moved to a tenure-track appointment in the Allen School in September 2018. Her curriculum vitae records promotion to associate professor in September 2022 and professor in September 2025.[2][3]

At the University of Washington, she holds the Torode Family Professorship in Computer Science & Engineering and leads H2Lab. The university also lists adjunct affiliations in Electrical & Computer Engineering and Linguistics. H2Lab's work spans language models, reasoning, retrieval, instruction following, evaluation, and applications involving scientific text.[1][5]

Her institutional appointments can be summarized as follows:

| Period | Institution | Appointment |
| --- | --- | --- |
| 2011-2012 | Disney Research, Pittsburgh | Postdoctoral associate |
| 2012-2015 | University of Washington Electrical Engineering | Research scientist |
| 2015-2018 | University of Washington Electrical Engineering | Research assistant professor |
| 2018-2021 | Allen Institute for AI | AI Fellow |
| 2018-2022 | University of Washington Allen School | Assistant professor |
| 2021-2026 | Allen Institute for AI | Senior director of research |
| 2022-2025 | University of Washington Allen School | Associate professor |
| 2025-present | University of Washington Allen School | Professor |
| 2026-present | Microsoft AI | Member of the Superintelligence team |

Hajishirzi joined Ai2 as an AI Fellow in 2018 and became a senior director of research in October 2021. Her June 2025 CV and some institutional biography pages still describe that appointment as current. Microsoft recruited her in 2026, however, and a Microsoft spokesperson later confirmed that she and other former Ai2 researchers had joined the company's Superintelligence team to work on its research program and model post-training. No more specific Microsoft title has been publicly confirmed.[2][4]

## Research

### Probabilistic reasoning and multimodal understanding

Hajishirzi's doctoral research treated reasoning in changing, partly observed environments. That work produced papers at AAAI, UAI, IJCAI, and the International Conference on Principles of Knowledge Representation and Reasoning. It also supplied a technical basis for later projects in which a system had to connect language to events, objects, or diagrams rather than classify text in isolation.[2][6]

Her early University of Washington work included systems that combined language with visual structure. The 2015 EMNLP paper *Solving Geometry Problems: Combining Text and Diagram Interpretation*, written with Minjoon Seo, Ali Farhadi, Oren Etzioni, and Clint Malcolm, addressed geometry questions containing both a diagram and a natural-language problem statement. The system constructed compatible interpretations of the two sources before solving for the requested quantity.[8]

This line of work continued in research on diagrams, visual knowledge extraction, semantic segmentation, and language-and-vision reasoning. It accounts for the [computer vision](https://aiwiki.ai/wiki/computer_vision) component of Hajishirzi's publication record, while her main academic appointment and laboratory have remained centered on AI and language processing.[2][5]

### Machine comprehension and question answering

A recurring subject in Hajishirzi's research is [question answering](https://aiwiki.ai/wiki/question_answering), especially systems that must find and combine evidence rather than match a question to a short memorized response. The Bi-Directional Attention Flow network, or BiDAF, was introduced by Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hajishirzi at ICLR 2017. BiDAF represented a passage and a query at several levels and used attention in both directions to form a query-aware passage representation. The paper reported results on [SQuAD](https://aiwiki.ai/wiki/squad) and the CNN/DailyMail cloze test.[9]

Later work in her group examined open-domain retrieval, multi-hop reasoning, scientific claim verification, numerical reasoning, and information extraction. These projects treated access to external evidence as part of the model's task. They also connected question answering to the construction and use of scientific knowledge bases.[1][5]

The ACL 2022 paper *Generated Knowledge Prompting for Commonsense Reasoning* tested a different route to external knowledge. Its method first prompted a language model to generate statements related to a question, then supplied those statements as extra context for answering it. The reported experiments covered numerical, general, and scientific commonsense tasks.[10]

### Instruction following, retrieval, and self-evaluation

Hajishirzi coauthored several projects on adapting a [large language model](https://aiwiki.ai/wiki/large_language_model) to new tasks through natural-language instructions. *Self-Instruct*, published at ACL 2023, generated instructions, inputs, and outputs from a language model, filtered invalid or overly similar examples, and then used the resulting data for [instruction tuning](https://aiwiki.ai/wiki/instruction_tuning). The project presented synthetic instruction generation as a way to reduce reliance on manually written task examples.[11]

Her work also addressed when a model should consult an external source. *Self-RAG*, presented at ICLR 2024, trained a model to retrieve passages on demand and to assess both the retrieved evidence and its own generated text through special reflection tokens. The method joined retrieval, generation, evidence assessment, and response control in one model rather than applying retrieval as an identical preliminary step for every request.[12]

These projects reflect a broader research program on the behavior of language models after pretraining. Topics in that program include data selection, factuality, model merging, reward-model evaluation, supervised fine-tuning, preference optimization, and [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) for instruction following.[1][15]

### OLMo and open model development

At Ai2, Hajishirzi co-led the OLMo project with researchers working on data, pretraining, evaluation, and post-training. The first OLMo release in February 2024 included model weights together with the training corpus, training and evaluation code, intermediate checkpoints, and logs. The release was designed so that researchers could inspect and modify more of the development process than is possible with a weights-only model release.[13][14]

The peer-reviewed OLMo paper, published in the ACL 2024 long-paper proceedings, lists Hajishirzi as a senior coauthor affiliated with both Ai2 and the University of Washington. It describes the model, the Dolma training corpus, training setup, evaluation framework, and the artifacts released for reuse. ACL 2024 gave the paper a Best Theme Paper Award. The related Dolma corpus paper, on which Hajishirzi was also a coauthor and adviser, received a Best Resource Paper Award.[14][22]

The project continued with OLMo 2 and OLMo 3, which changed the training data, architecture, model sizes, post-training stages, and released checkpoints while keeping the broader emphasis on inspectable development. Hajishirzi's official biography continued to identify her as an OLMo co-lead through 2025.[1]

### Tülu and post-training

Tülu is the companion line of work on turning a pretrained model into an instruction-following model. Hajishirzi co-led the project during her time at Ai2. Tülu 3 documented a pipeline that included curated instruction data, supervised fine-tuning, preference optimization, reinforcement learning with verifiable rewards, decontamination, and evaluation. Ai2 released the training mixtures, code, evaluation tools, and model checkpoints along with the report.[15][16]

The Tülu 3 paper was first released in 2024 and appeared at the Conference on Language Modeling in 2025. Its author list includes Hajishirzi along with researchers who worked on data curation, training algorithms, evaluation, and infrastructure. The project is relevant to her research on instruction following because it studies the combined effect of several post-training stages instead of treating fine-tuning as a single operation.[16]

### Vision-language models and scientific infrastructure

Hajishirzi was also a coauthor of *Molmo and PixMo: Open Weights and Open Data for State-of-the-Art Vision-Language Models*, published in the CVPR 2025 proceedings. [Molmo](https://aiwiki.ai/wiki/molmo) is a family of [vision-language models](https://aiwiki.ai/wiki/vision_language_model), while PixMo is a collection of image-caption, visual question-answering, document, and pointing data. The paper emphasized collecting training data without using another vision-language model to generate image descriptions, and it released model weights, datasets, and source code.[17]

In August 2025, the U.S. National Science Foundation and NVIDIA announced a combined $152 million in support for the Open Multimodal AI Infrastructure to Accelerate Science project, known as OMAI. The official project page lists Hajishirzi as a co-principal investigator from the University of Washington. OMAI is intended to produce open models and infrastructure for scientific work, building on the OLMo and Molmo projects. The page identifies NSF Award No. 2413244 and continued to list Hajishirzi as a co-principal investigator after her move to Microsoft.[18]

## Selected works

| Year | Work | Venue or form | Main subject |
| --- | --- | --- | --- |
| 2011 | *Action-centered Reasoning for Probabilistic Dynamic Systems* | Ph.D. dissertation, University of Illinois | Probabilistic reasoning about actions and dynamic systems[6] |
| 2012 | *Using Group History to Identify Character-Directed Utterances in Multi-Child Interactions* | SIGDIAL | Multimodal addressee identification in group interaction[7] |
| 2015 | *Solving Geometry Problems: Combining Text and Diagram Interpretation* | EMNLP | Joint interpretation of language and diagrams[8] |
| 2017 | *Bidirectional Attention Flow for Machine Comprehension* | ICLR | Passage-question attention for machine comprehension[9] |
| 2022 | *Generated Knowledge Prompting for Commonsense Reasoning* | ACL | Language-model-generated knowledge as reasoning context[10] |
| 2023 | *Self-Instruct: Aligning Language Models with Self-Generated Instructions* | ACL | Synthetic data for instruction following[11] |
| 2024 | *Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection* | ICLR | Adaptive retrieval and output assessment[12] |
| 2024 | *OLMo: Accelerating the Science of Language Models* | ACL | Open language-model data, code, checkpoints, and evaluation[14] |
| 2025 | *Tülu 3: Pushing Frontiers in Open Language Model Post-Training* | COLM | Open post-training recipes and evaluation[16] |
| 2025 | *Molmo and PixMo: Open Weights and Open Data for State-of-the-Art Vision-Language Models* | CVPR | Open multimodal models and training data[17] |

## Awards and professional recognition

Hajishirzi received a Sloan Research Fellowship in computer science in 2020. The University of Washington announcement identifies her work at the time as spanning representation learning, question answering, knowledge graphs, dialogue, and knowledge extraction.[19]

In 2021, the National Science Foundation awarded her a CAREER grant for research on knowledge-rich neural text comprehension and reasoning. The project focused on systems that could find evidence in varied online sources, integrate explicit and implicit information, and use the result to answer questions or verify claims.[20]

The University of Illinois computer science program gave Hajishirzi its 2023 Early Career Academic Achievement Alumni Award. In 2022 she received the Torode Family Career Development Professorship at Washington, and the Allen School now lists her as the Torode Family Professor in Computer Science & Engineering.[2][5][21]

The Association for Computational Linguistics selected Hajishirzi as an ACL Fellow in 2025. This recognition followed the two ACL 2024 paper awards associated with OLMo and Dolma. Her earlier SIGDIAL best paper award dates to 2012, when the recognized work concerned language interaction between groups of children and an animated character.[7][22][23]

## References

1. Hajishirzi, Hanna. "Hanna Hajishirzi." Official homepage, University of Washington. https://hannaneh.ai/
2. Hajishirzi, Hannaneh. "Curriculum Vitae." Last updated June 2025. https://hannaneh.ai/CV_Hanna.pdf
3. University of Washington Paul G. Allen School of Computer Science & Engineering. "Allen School strengthens its leadership in AI with the arrival of Hannaneh Hajishirzi." July 9, 2018. https://news.cs.washington.edu/2018/07/09/allen-school-strengthens-its-leadership-in-ai-with-the-arrival-of-hannaneh-hajishirzi/
4. Bishop, Todd. "Microsoft adds more former Ai2 researchers, bolstering its Superintelligence team." GeekWire, May 15, 2026. https://www.geekwire.com/2026/microsoft-adds-more-former-ai2-researchers-to-its-superintelligence-team/
5. University of Washington Paul G. Allen School of Computer Science & Engineering. "Hannaneh Hajishirzi, Faculty." https://www.cs.washington.edu/people/faculty/hannaneh-hajishirzi/
6. Hajishirzi, Hannaneh. *Action-centered Reasoning for Probabilistic Dynamic Systems*. Ph.D. dissertation, University of Illinois at Urbana-Champaign, 2011. https://www.ideals.illinois.edu/items/29911
7. Hajishirzi, Hannaneh, Jill F. Lehman, and Jessica K. Hodgins. "Using Group History to Identify Character-Directed Utterances in Multi-Child Interactions." *Proceedings of SIGDIAL 2012*, pp. 207-216. https://aclanthology.org/W12-1629/
8. Seo, Minjoon, Hannaneh Hajishirzi, Ali Farhadi, Oren Etzioni, and Clint Malcolm. "Solving Geometry Problems: Combining Text and Diagram Interpretation." *Proceedings of EMNLP 2015*, pp. 1466-1476. https://aclanthology.org/anthology-files/pdf/D/D15/D15-1171.pdf
9. Seo, Minjoon, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. "Bidirectional Attention Flow for Machine Comprehension." *ICLR 2017*. https://openreview.net/forum?id=HJ0UKP9ge
10. Liu, Jiacheng, et al. "Generated Knowledge Prompting for Commonsense Reasoning." *Proceedings of ACL 2022*, pp. 3154-3169. https://aclanthology.org/2022.acl-long.225/
11. Wang, Yizhong, et al. "Self-Instruct: Aligning Language Models with Self-Generated Instructions." *Proceedings of ACL 2023*, pp. 13484-13508. https://aclanthology.org/2023.acl-long.754/
12. Asai, Akari, Zeqiu Wu, Yizhong Wang, Avi Sil, and Hannaneh Hajishirzi. "Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection." *ICLR 2024*. https://proceedings.iclr.cc/paper_files/paper/2024/hash/25f7be9694d7b32d5cc670927b8091e1-Abstract-Conference.html
13. Allen Institute for AI. "OLMo: Open Language Model." February 1, 2024. https://allenai.org/blog/olmo-open-language-model-87ccfc95f580
14. Groeneveld, Dirk, et al. "OLMo: Accelerating the Science of Language Models." *Proceedings of ACL 2024*, pp. 15789-15809. https://aclanthology.org/2024.acl-long.841/
15. Allen Institute for AI. "Tülu 3: The next era in open post-training." November 21, 2024. https://allenai.org/blog/tulu-3-technical
16. Lambert, Nathan, et al. "Tülu 3: Pushing Frontiers in Open Language Model Post-Training." *Second Conference on Language Modeling*, 2025. https://arxiv.org/abs/2411.15124
17. Deitke, Matt, et al. "Molmo and PixMo: Open Weights and Open Data for State-of-the-Art Vision-Language Models." *Proceedings of CVPR 2025*, pp. 91-104. https://openaccess.thecvf.com/content/CVPR2025/html/Deitke_Molmo_and_PixMo_Open_Weights_and_Open_Data_for_State-of-the-Art_CVPR_2025_paper.html
18. Allen Institute for AI. "NSF Mid-Scale RI-2: Open Multimodal AI Infrastructure to Accelerate Science." https://allenai.org/omai
19. University of Washington Paul G. Allen School of Computer Science & Engineering. "Hannaneh Hajishirzi and Yin Tat Lee named 2020 Sloan Research Fellows." February 12, 2020. https://news.cs.washington.edu/2020/02/12/hannaneh-hajishirzi-and-yin-tat-lee-named-2020-sloan-research-fellows/
20. University of Washington Paul G. Allen School of Computer Science & Engineering. "CAREER Award-winning faculty at the Allen School advance leadership and innovation in software testing, machine learning equity, and natural language understanding." August 13, 2021. https://news.cs.washington.edu/2021/08/13/career-award-winning-faculty-at-the-allen-school-advance-leadership-and-innovation-in-software-testing-machine-learning-equity-and-natural-language-understanding/
21. University of Illinois Siebel School of Computing and Data Science. "Hanna Hajishirzi, 2023 Early Career Academic Achievement Alumni Award." https://siebelschool.illinois.edu/about/awards/alumni-awards/alumni-awards-past-recipients/hajishirzi
22. Association for Computational Linguistics. "Best Paper Awards, ACL 2024." https://2024.aclweb.org/program/best_papers/
23. Association for Computational Linguistics. "ACL Newsletter, 2025." https://www.aclweb.org/portal/content/acl-newsletter-2025