Federated evaluation of agent-trajectory evaluators

Develop methods for federating evaluators of agent trajectories so that clients can collaboratively improve trajectory assessment while preserving private outcomes and accommodating heterogeneous client utilities.

Background

Federated Agent Optimization treats reward and evaluation mechanisms as one of the agent components that may be collaboratively optimized. Clients possess private real-world outcomes and human judgments, while a coordinator may have stronger general-purpose evaluators but lacks access to those outcomes.

The paper distinguishes existing federated preference and reward-model methods for language-model alignment from the unresolved problem of federating evaluators specifically for agent trajectories. Such evaluators must handle sparse, delayed, and heterogeneous feedback while limiting disclosure of individual case outcomes.