The Horcrux: Mechanistically Interpretable Task Decomposition for Detecting and Mitigating Reward Hacking in Embodied AI Systems

TL;DR

This paper introduces MITD, a hierarchical transformer architecture that decomposes tasks into interpretable subtasks to detect and reduce reward hacking in embodied AI, outperforming traditional monitoring methods.

Contribution

The paper presents a novel mechanistically interpretable task decomposition method using hierarchical transformers to address reward hacking in embodied AI systems.

Findings

Decomposition depths of 12-25 steps reduce reward hacking by 34%.

MITD provides diagnostic visualizations like Attention Waterfall Diagrams.

Mechanistically grounded decomposition outperforms post-hoc behavioral monitoring.

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.