Mine Training Data from Noisy UI Interaction Logs

Asked at OpenAI · System Design for ML · Difficulty: Hard · Last asked Nov 2, 2025

Design a data and training system that converts a large, noisy collection of screen observations and user actions into useful supervision for a multimodal UI-navigation agent. The corpus mixes successful, failed, accidental, and incomplete trajectories, so the system must estimate quality without discarding informative failures or reinforcing its own model errors.

Loading the problem…