- By Eric Horvitz, Chief Scientific Officer, and Saurabh Tiwary, Corporate VP and Technical Fellow, Microsoft Turing 

The AI research community is focusing attention on large language models (LLMs) given their impressive performance with difficult tasks and value in useful applications. The capabilities of the models have stimulated many research questions: How do these models work and what do they learn? How can we leverage their potential for new innovations and discoveries? How can we mitigate potentially harmful behaviors? These and many other questions are inspiring new and challenging directions for AI research and underscoring the need for different perspectives and proficiencies. 
 
To explore what’s ahead for LLMs and discuss the ways industry, academia, and government could work together to advance understanding of these models, we organized a panel titled, “Towards a Healthy Research Ecosystem for Large Language Models.” We were joined by Ahmed Awadallah from Microsoft Research, Erwin Gianchandani from the National Science Foundation, and Percy Liang from Stanford University who each brought fascinating insights and ideas for consideration. A key part of the discussion is on the need to expand access to large language models. Building and experimenting with the largest models requires a great deal of data and computing resources, which are often beyond the reach of university-based teams. This challenge was our motivation for launching the Microsoft Turing Academic Program (MS-TAP). Since 2021, we have sought to provide leading academic teams with access to some of the world's largest language models. The program reflects our belief in the importance of having diverse and talented teams from academia working with these models.

MS-TAP has supported multiple in-depth collaborations with partner universities. With deep engagement from researchers and domain experts in Microsoft Research, Microsoft Turing, and the Office of the Chief Scientific Officer, we work to better understand model behavior, identify novel applications, explore potential risks, develop mitigations, and improve future models. Participants receive unprecedented access to our 530B parameter Natural Language Generation model (T-NLGv2), Natural Language Representation model (T-NLRv5), and 2.5B parameter Universal Image Language Representation model (T-UILRv2) and Azure compute resources to run experiments and evaluations.
 
MS-TAP Phase 1 involved collaborations with six universities on five projects (see below) and included numerous highlights like these papers: Was it “said” or was it “claimed”? How linguistic bias affects generative language models (Brown University) and Invariant Language Modeling (École Polytechnique Fédérale de Lausanne) published at the EMNLP 2022 conference. More information including contributors and ongoing updates is available on the MS-TAP Phase 1 Collaborations page.
 

• University of California, Berkeley and University of California, San Francisco: Leveraging large language models for transfer learning in medical notes
• Brown University: The extent to which large language models exacerbate bias when given different types of biased and unbiased inputs
• École Polytechnique Fédérale de Lausanne (EPFL): Enhancing the robustness of massive language models via invariant representation learning
• Georgia Tech: Analyzing and using large pretrained language models for societal good
• University of Washington: Analyzing toxicity, factuality and memory

The second phase is currently underway and focuses on larger and more complex models. We are collaborating with seven universities on nine projects. Additional details about each effort including contributors are available on the MS-TAP Phase 2 Collaborations page. We will share links to forthcoming papers at the close of Phase 2.
 

• Carnegie Mellon University: Large language models for dialog evaluation
• Carnegie Mellon University: Learning instructible visuo-motor agents through multimodal interactive teaching
• École Polytechnique Fédérale de Lausanne (EPFL): Impact of Decoding Strategies for LLMs
• Harvard University: Transferring word representations to the electronic health records with disparity
• Harvard University: Improving the reasoning ability of large pretrained models by instructional scaffolding
• Massachusetts Institute of Technology: Speeding up training and fine-tuning for large-scale NLP models (Paper: SmoothQuant: Accurate and Efficient Post Training Quantization for Large Language Models)
• Mila – Quebec Artificial Intelligence Institute: Reducing the impact of Summaries generated by LLMs
• University of Michigan: Enabling transparency and interpretability in Turing natural language representation models
• Stanford University: A multi-faceted benchmark for large language models (Paper: Holistic Evaluation of Language Models; Resource: Language Models are Changing AI. We Need to Understand Them)

We look forward to continuing our collaborations with our academic partners and to welcoming new ones in future phases.

Academic labs interested in participating in future phases of the program should send email to turingacademic@microsoft.com

Resources
Microsoft Turing Academic Program
Microsoft Turing
Microsoft Responsible AI
Microsoft Research Summit 2022