Percy Liang

Overview

Percy Liang is a faculty member in computer science at Stanford University. He directs the Center for Research on Foundation Models (CRFM) and is a Senior Fellow at Stanford HAI.^[1][2] The two roles point at the same problem from different sides: how very large, general-purpose models are built, and how people outside the group that built them can use, inspect and compare them.

Education

He holds a B.S. from MIT (2004), an M.Eng. from MIT (2005) and a Ph.D. from UC Berkeley (2011).^[1][2] Undergraduate and master's training at one institution followed by doctoral work at another is a common path for computer science researchers in the United States, and it means roughly a decade of formal training stands behind the faculty appointment.

Current focus

His Stanford HAI profile states the focus directly: making foundation models, especially language models, more accessible through open-source work and more understandable through rigorous benchmarking.^[2]

Both halves deserve unpacking. A foundation model is a large model trained on broad data and then adapted to many downstream tasks, rather than built for a single purpose. Open-source release means publishing code and weights so that other researchers, students and companies can run, examine and modify the model instead of only reaching it through someone else's interface. Benchmarking means testing models on shared, standardized sets of tasks, so that claims about which model is better rest on the same measurements rather than on anecdotes. The two goals reinforce each other: a model that outsiders can download is also a model outsiders can check.

Earlier research

His past work covers robustness, interpretability, human interaction, learning theory, grounding, semantics and reasoning.^[2] These are broad labels, but they share a theme — understanding what a model has actually learned and where it fails. Robustness asks whether a system keeps working when the data shifts; interpretability asks why it produced a particular output; grounding asks how words connect to things in the world; semantics and reasoning ask what a model's sentences mean and whether it can chain steps together; human interaction asks how people steer the system; learning theory supplies the mathematics for when learning from data can be expected to generalize.

He also promotes reproducibility through CodaLab Worksheets, a platform for running computational experiments in a form that others can rerun and share, so that published results become something another lab can verify rather than take on trust.^[2]

Recognition

Liang has received the Presidential Early Career Award for Scientists and Engineers (2019), the IJCAI Computers and Thought Award (2016) and an NSF CAREER Award (2016).^[1][2] He also received a Sloan Research Fellowship (2015) and a Microsoft Research Faculty Fellowship (2014).^[1][2] Several of these are early-career honours: PECASE is given by the US government to young scientists and engineers, the Sloan fellowship and NSF CAREER support researchers at the start of independent careers, and the Computers and Thought Award is presented at IJCAI to a young AI researcher.

His Stanford CS biography also lists paper awards at ACL, EMNLP, ICML, COLT, ISMIR, CHI, UIST and RSS.^[1] The spread is wide: ACL and EMNLP are natural language processing venues, ICML and COLT are machine learning and learning theory venues, CHI and UIST are human-computer interaction venues, RSS is a robotics venue, and ISMIR is a music information retrieval venue. Recognition across that range suggests a research programme that moves between statistical foundations, language, interaction and physical systems.

Place in the field

Directing a centre devoted to foundation models while treating open release and shared benchmarks as core research goals means his influence appears not only in individual papers but in how the field measures itself. Benchmarks and open toolkits are infrastructure: less conspicuous than a new system, but they shape what counts as progress. Because the models in question change quickly, a common way of measuring them is what allows results to be compared at all, and a common way of releasing them is what lets people outside a small set of labs study them.