The Horcrux: Mechanistically Interpretable Task Decomposition for Detecting and Mitigating Reward Hacking in Embodied AI Systems
Subramanyam Sahoo, Jared Junkin
TL;DR
Reward hacking presents a major safety risk for embodied AI. MITD introduces a hierarchical Planner–Coordinator–Executor architecture that decomposes tasks into interpretable subtasks and provides visual diagnostics to trace internal reasoning. In experiments with 1,000 HH-RLHF samples, a decomposition depth in the 12–25 step range minimizes reward hacking across multiple failure modes, outperforming post-hoc monitoring approaches. The work demonstrates that mechanistically grounded task decomposition, coupled with interpretability visualizations, offers actionable insights for detecting and mitigating misalignment in complex agents, with potential for scalable safety oversight through targeted interventions.
Abstract
Embodied AI agents exploit reward signal flaws through reward hacking, achieving high proxy scores while failing true objectives. We introduce Mechanistically Interpretable Task Decomposition (MITD), a hierarchical transformer architecture with Planner, Coordinator, and Executor modules that detects and mitigates reward hacking. MITD decomposes tasks into interpretable subtasks while generating diagnostic visualizations including Attention Waterfall Diagrams and Neural Pathway Flow Charts. Experiments on 1,000 HH-RLHF samples reveal that decomposition depths of 12 to 25 steps reduce reward hacking frequency by 34 percent across four failure modes. We present new paradigms showing that mechanistically grounded decomposition offers a more effective way to detect reward hacking than post-hoc behavioral monitoring.