🧑‍🔬 Trevor Cai

Trevor Cai works on large language models: how they are trained and how they are tested for safety. The papers here, both from 2022, fall under Language models & pretraining and Bias, fairness & safety. They share a focus on transformer systems at very large scale, studied through extensive training runs and through automated evaluation of model behaviour.

The most cited of Cai's papers is Training Compute-Optimal Large Language Models. It asks how a fixed compute budget should be divided between model size and training data. The authors trained a large family of models of varying sizes on varying numbers of tokens and estimated the best balance in three separate ways. Each method pointed to the same answer: parameters and training tokens should grow in step. By that measure, many prominent models of the period were oversized and undertrained. The second paper, Red Teaming Language Models with Language Models, replaces hand-written safety probes with an automated loop. One language model generates test questions, the target model responds, and a classifier flags harmful replies. The method was applied to a dialogue-prompted Gopher chatbot and compared several strategies for generating test cases, which differ in how diverse their questions are and how often they expose failures.

Both papers date from the same year, so the recent direction is the same pairing of efficient scaling and safety testing. Bylines from 2022 connect the work to New York University. Co-authors on these papers include Ethan Perez, Geoffrey Irving, Arthur Mensch, Amelia Glaese and Johannes Welbl.

Trevor Cai has 2 papers in our library, in 2022, most of them on Language models & pretraining, Bias, fairness & safety; often with Aidan Clark, Amelia Glaese, Arthur Mensch.