Dual-Path Anti-Drift for Reliable Long-Horizon LLM Agents: A Formal and Mechanistic Simulation Study

Abstract

As AI agents take on longer and more consequential tasks, a central reliability problem becomes harder to ignore: a system can appear successful while drifting away from verified evidence. A fluent answer, a completed workflow, or a confident final claim is not necessarily a correct one. This paper introduces Dual-Path Anti-Drift (DPAD), a framework that separates problem solving from critical assessment. A primary path pursues task completion, while an isolated critic searches for shortcuts, unsupported confidence, and missing evidence under a separate objective function. A deterministic coordinator then governs how criticism can affect the final decision. Through formal analysis and a transparent mechanistic simulation, the study examines the trade-offs among false-success reduction, task completion, abstention, and computational cost. Rather than claiming a finished solution, DPAD offers a testable architecture and a clear path toward more reliable long-horizon AI agents.

Keywords