A research paper introduces Task Model Induction (TMI), a method that discovers latent tasks from unconstrained computer-use traces and induces hierarchical task models. On controlled human and agent trajectories, TMI recovers interleaved tasks with 0.974 agreement against ground-truth groupings and reconstructs 74.9% of observed execution.
The paper addresses the challenge of deriving symbolic, auditable, and reusable task models from naturalistic computer-use traces, which are passively recorded screenshots and input actions. Existing methods assume a given task or single workflow and produce step-level summaries rather than structured models. TMI discovers latent tasks in unconstrained traces, disentangles concurrent activity, and induces a task model pairing a hierarchical objective model with a procedure model of control flow. Intrinsic evaluation shows high agreement with ground-truth task groupings and partial reconstruction of execution.
TMI operates on low-level event traces without task labels, using latent task discovery to handle multi-threaded, interleaved work. It produces structured models with recursive goal decomposition and control flow, enabling auditability and reuse. The reported 0.974 agreement on task grouping and 74.9% execution reconstruction suggest robust performance on controlled trajectories, but generalization to noisy real-world traces remains to be validated.
As computer-use agents enter real work, organizations need auditable and reusable knowledge of how tasks are performed. TMI could enable automatic documentation of workflows, compliance auditing, and agent training from observed human behavior. Adoption will depend on robustness to real-world trace noise and integration with existing process mining or RPA tools.
Potential value lies in automating process discovery, reducing manual workflow documentation, and enabling agents to learn from human demonstrations. This could lower costs for enterprise automation and improve agent reliability. However, value is unproven until validated outside controlled settings.
Next signals include peer-reviewed publication, open-source code release, and evaluation on real-world enterprise traces. If TMI scales to noisy data, it may influence agent training pipelines and enterprise process intelligence. Watch for collaborations with RPA vendors or agent platforms.