▶ Watch ↗17:52
Conference affiliation: Netflix
Cameron R. Wolfe is a staff research scientist on Netflix’s Globalization team, working on multimodal large language models. His research ranges from reducing the computation needed to train and run neural networks to evaluating outputs whose quality cannot be captured by a single correct answer. At Netflix, that includes judging promotional synopses and translation quality. He also writes Deep (Learning) Focus, explaining AI research through its mathematics, implementation details, and practical implications.
Cameron studied computer science at the University of Texas at Austin, where he researched genetic algorithms and evolutionary computation with the Neural Networks Research Group. He earned his computer science Ph.D. at Rice University in 2023, advised by Anastasios Kyrillidis. Alongside his academic research, he held a year-round data-science internship at Salesforce, worked in research at Alegion, and became director of AI at Rebuy Engine. He joined Netflix as a senior research scientist in 2024 and became a staff research scientist in 2026.
At Rice, Cameron investigated how much of a neural network’s computation could be removed or reorganized while preserving its behavior. With Kyrillidis, he developed i-SpaSP, a structured pruning algorithm inspired by sparse signal recovery. It repeatedly identifies groups of neurons or filters that help explain the difference between a full network’s output and a smaller network’s output, then selects which groups to retain. The work combined experiments on image and language models with theoretical guarantees about pruning error.
With John Chen and Kyrillidis, Cameron also developed REX, a learning-rate schedule evaluated across different training budgets. The researchers examined how schedules behaved when training stopped after a small fraction of the usual computation; their proposed schedule required no additional computation, storage, or hyperparameters. In a 2024 interview with Rice, Cameron connected his doctoral work on independent subnetwork training, pruning, and quantization to the demands of training and hosting large networks at Netflix.
At Netflix, Cameron worked on tools for promotional writing, including the short descriptions that help viewers decide what to watch. A synopsis can be factually accurate yet unclear, tonally wrong, or misleading about the experience a show offers. Evaluation therefore requires a definition of quality that goes beyond checking facts.
With Gabriela Alessio and Cameron Taylor, he developed an LLM-as-a-Judge system for Netflix synopses. Creative experts helped define the standards, and disagreements between their ratings drove revisions to the scoring guidelines. The team then calibrated automated judges against expert judgments.
The team found that one prompt covering every quality criterion performed poorly. Separate judges handled individual criteria, and factuality checks were divided further: verifying plot information required different supporting context from checking cast or awards information. Narrowing each evaluator’s task made the judgments more manageable. Judge scores also predicted differences in viewing behavior, including abandonment. Those associations supported the system’s practical usefulness without establishing that changing a judge score itself caused a change in behavior.
In his AI Engineer talk on subjective evaluation, Cameron explained why agreement with experts is only part of the problem. A model and its automated judge can both produce different results across repeated runs. He advocates power analysis to estimate how many examples an evaluation needs, based on the improvement a team wants to detect and the variation in its results. Human agreement provides a baseline for judging an automated evaluator; production experiments test whether measured improvements matter in use. Failures found after deployment can then become examples in subsequent evaluations.
In TQLite, co-authored with Bhavin Jawade in 2026, Cameron addressed the cost of translation-quality evaluation. The method uses a jury of large reasoning models to create and curate synthetic training examples, then distills their judgments into smaller language models. The aim is to approach the stronger evaluators’ performance at a lower cost, making evaluation more affordable at scale.
Cameron’s explanatory work includes both implementations and long-form articles:
▶ Watch ↗17:52
Loading more talks and highlights…