Open sourced work:
API for computer use(prava-sdk, ca-server)
Models: Grounding+planning hardness for OSWorld. Privately I used codex+extensive planning to benchmark max but OSWorld scores were brittle.
Research:
a) Edited the vision encoder such that we would show the latest "compiled" screenshot instead of having to re-encode the whole thing again, which worked but was slightly slower
b) re-implemented MSFT's GUI-Actor which essentially trains a small attention head to try predict which patch/region instead of coordinate grounding
API: CUA Control API docs
Control API library implementation
Models: prava-fc doesnt include latest models which use codex(for strong tool calling) + qwen3 for good grounding + codex to write mini bash scripts to enhance computer control
Research: Differential Vision Encoding for CUA
