Job Information
Microsoft Corporation Principal Applied Scientist in Redmond, Washington
We are seeking a Principal Applied Scientist to join our organization and help improve steerability and control Large Language Models (LLMs) and other AI systems. Our team currently develops Guidance, a fully open-source project that enables developers to control language models more precisely and efficiently with constrained decoding.
As a Principal Applied Scientist, you will play a crucial role in advancing the research frontier of constrained decoding. If you’re excited about links between formal grammars and generative AI, deeply understanding and optimizing LLM inference, enabling more responsible AI without finetuning and Reinforcement Learning from Human Feedback (RLHF), and/or exploring fundamental changes to the “text-in, text-out” application programming interface (API), we’d love to hear from you. Our team offers a vibrant environment for cutting-edge, multidisciplinary research. We have a long track record of open-source code and open publication policies, and you’ll have the opportunity to collaborate with world-leading experts across Microsoft and top academic institutions across the world.
Microsoft’s mission is to empower every person and every organization on the planet to achieve more. As employees we come together with a growth mindset, innovate to empower others, and collaborate to realize our shared goals. Each day we build on our values of respect, integrity, and accountability to create a culture of inclusion where everyone can thrive at work and beyond. In alignment with our Microsoft values, we are committed to cultivating an inclusive work environment for all employees to positively impact our culture every day.
Responsibilities
Develop new constrained decoding research techniques for increasing LLM inference quality and/or efficiency. Example areas of interest include speculative execution, new decoding strategies (e.g. extensions to beam search), “classifier in the loop” decoding for responsible AI, improving AI planning, and explorations of attention-masking based constraints.
Re-imagine the use and construction of context-free grammars (CFG) and beyond to fit Generative AI. Examples of improvements here include better tools for constructing formal grammars, extensions to Earley parsing, and efficient batch processing for constrained generation. Consideration of how these techniques are presented to developers – who may not be well versed in grammars and constrained generation -- in an intuitive, idiomatic programming syntax is also top of mind.
Design principled evaluation frameworks and benchmarks for measuring the effects of constrained decoding on a model. Some areas of interest to study carefully include efficiency (token throughput and latency), generation quality, and impacts of constrained decoding on AI safety.
Publish your research in top AI conferences and contribute your research advances to the guidance open-source project.
Qualifications
Required/Minimum Qualifications
Bachelor's Degree in Statistics, Econometrics, Computer Science, Electrical or Computer Engineering, or related field AND 6+ years related experience (e.g., statistics, predictive analytics, research)
OR Master's Degree in Statistics, Econometrics, Computer Science, Electrical or Computer Engineering, or related field AND 4+ years related experience (e.g., statistics, predictive analytics, research)
OR Doctorate in Statistics, Econometrics, Computer Science, Electrical or Computer Engineering, or related field AND 3+ years related experience (e.g., statistics, predictive analytics, research)
OR equivalent experience.
Preferred Qualifications:
· Master's Degree in Statistics, Econometrics, Computer Science, Electrical or Computer Engineering, or related field AND 9+ years related experience (e.g., statistics, predictive analytics, research)
OR Doctorate in Statistics, Econometrics, Computer Science, Electrical or Computer Engineering, or related field AND 6+ years related experience (e.g., statistics, predictive analytics, research)
OR equivalent experience.
Expertise on one of the following preferred:
Deep familiarity with transformer-based model inference, including batch processing paradigms for hosted models
Expertise in context-free grammar specification and parsing
Experience with constrained decoding paradigms (regex-based constraints, grammar based constraints, JSON mode, function calling, etc.)
Contribution history to open-source projects, especially in the LLM/AI space
Publication history in top AI conferences (e.g. NeurIPS, ICML, ICLR, EMNLP, ACL, AAAI)
Deep familiarity with Python programming paradigms and modern LLM APIs
Effective communication skills and desire to collaborate in a multi-disciplinary team
Familiarity with Guidance
Applied Sciences IC5 - The typical base pay range for this role across the U.S. is USD $133,600 - $256,800 per year. There is a different range applicable to specific work locations, within the San Francisco Bay area and New York City metropolitan area, and the base pay range for this role in those locations is USD $173,200 - $282,200 per year.
Certain roles may be eligible for benefits and other compensation. Find additional benefits and pay information here: https://careers.microsoft.com/us/en/us-corporate-pay
Microsoft is an equal opportunity employer. Consistent with applicable law, all qualified applicants will receive consideration for employment without regard to age, ancestry, citizenship, color, family or medical care leave, gender identity or expression, genetic information, immigration status, marital status, medical condition, national origin, physical or mental disability, political affiliation, protected veteran or military status, race, ethnicity, religion, sex (including pregnancy), sexual orientation, or any other characteristic protected by applicable local laws, regulations and ordinances. If you need assistance and/or a reasonable accommodation due to a disability during the application process, read more about requesting accommodations (https://careers.microsoft.com/v2/global/en/accessibility.html) .