Research · October 20, 2026
How work moves across the web
Where people go next, where their time goes, and what each kind of role actually opens all day, measured from a large panel of real browsing sessions. Search any site or role.
Research
Benchmarks, evaluations, and methods from the team building AI for professional computer work. Questions, or a model you'd like measured? Write to research@composite.com.
Research · September 25, 2026
JEPA style Temporal Representations, Diffusion, and Dynamics Models for Browser Action Prediction
Experiments that led us to intent as the bottleneck: JEPA-inspired next-state prediction, diffusion-based action generation, and browser dynamics as supervision.
Research · September 21, 2026
Pushing the Capabilities of Computer Use Agents with Intent Detection
Early research on predicting a user's next browser action without an explicit instruction, and why inferring intent, not grounding to the page, turned out to be the bottleneck.
Benchmarking · July 24, 2026
Introducing Composite-Bench: the strongest open-weights model isn't Kimi K3
The first long-horizon, browser-based computer-use benchmark with certified optimality and verified compute. In our evaluation, GLM-5.2 lands 11 points behind a two-way Claude tie and beats every closed model we measured.