Skip to content

Instantly share code, notes, and snippets.

View yogyam's full-sized avatar

Yogya Mehrotra yogyam

View GitHub Profile

Task Proposal: UI Behavior Profiling

Core Idea

An agent observes a simulated user interacting with a web app across multiple sessions. Each user has a hidden behavioral profile — a fixed combination of traits like navigation style, error tendency, help-seeking behavior, and shortcut usage. The agent receives partial interaction logs and predicts the user's next actions. A learning agent should improve over sessions; a stateless baseline stays flat.

Instance Structure

Each instance: agent receives a partial session log → predicts next K actions → scored against ground truth.