Microsoft Introduces AdsWorldEngine: Self-Evolving Agent for Conversational Ads

_reachsumit · x · 2026-08-17

Core Function: AdsWorldEngine is an agentic framework for conversational advertising. It uses an "Opportunity Gate" to decide ad placement, an Orchestrator to generate commercial intents and call ad tools, and an Evaluator to score ads for offline optimization.

Technical Innovation: The paper proposes an iterative actor-tool training procedure. It first trains the Orchestrator via SFT and agentic RL, then uses high- and low-reward rollouts to construct preference data for training tools, creating a self-improving loop that learns both how to use and how to improve the tools.

Value: Addresses the challenge of inferring latent commercial intent from multi-turn dialogue while balancing ad utility with intrusiveness.

Original post →

More from coding & agent

coding & agent channel →