🧑‍🔬 Pang Wei Koh

Pang Wei Koh works on understanding and evaluating machine learning models. Their papers here take up two recurring questions: where a model's predictions come from, and how far its outputs can be trusted. The methods range from statistical tools that trace a prediction back to the training set to evaluation schemes that check generated text against an external knowledge source. The later work falls under text generation with large language models.

The most cited of these papers is Understanding Black-box Predictions via Influence Functions (2017), written with Percy Liang. Most explanation methods ask how a trained model arrives at an output. This paper asks instead which training examples were responsible for it. It adapts influence functions from robust statistics to estimate how the parameters and the loss on a test point would shift if one training example were up-weighted, removed or perturbed. It avoids retraining the model and never forms or inverts the Hessian explicitly, which makes the approach workable for models with many parameters. The second paper, FActScore (2023), deals with factual accuracy in long generated passages. It breaks a response into atomic facts, each carrying a single piece of information, and reports the share of them that a chosen knowledge source supports. The paper tests this on biographies of people, using Wikipedia as the reference, and compares several language models using expert annotations.

The move from the 2017 paper to the 2023 one shifts the focus from explaining individual predictions to measuring how reliable large language models are. Koh's earlier work was done at Stanford University, and the current affiliations are the University of Washington and the Allen Institute for Artificial Intelligence in Seattle. Co-authors on the factuality work include Sewon Min, Luke Zettlemoyer and Hannaneh Hajishirzi.

Pang Wei Koh has 2 papers in our library, in 2017–2023, most of them on Text generation; often with Hannaneh Hajishirzi, Kalpesh Krishna, Luke Zettlemoyer.