PLOT: Progressive Localization via Optimal Transport in Neural Causal Abstraction

Researchers introduce PLOT (Progressive Localization via Optimal Transport), a new framework for mechanistic interpretability that efficiently identifies causal variables in neural networks through optimal transport coupling rather than computationally expensive searches. The method significantly speeds up causal abstraction analysis while maintaining competitive accuracy, offering practical advantages for large-scale AI interpretability research.

PLOT addresses a fundamental computational bottleneck in mechanistic interpretability research. Traditional approaches like distributed alignment search (DAS) require exhaustive searching across candidate neural sites to locate relevant causal variables, creating scalability challenges as models grow larger. By leveraging optimal transport theory, PLOT reformulates this localization problem geometrically, matching the output effects of abstract interventions with candidate neural sites through coupling analysis rather than brute-force enumeration.

The research builds on established causal abstraction frameworks but introduces a progressive refinement strategy particularly valuable for modern deep learning systems. Starting with coarse-grained sites (tokens, timesteps, layers), PLOT iteratively identifies finer-grained neural correlates, dramatically reducing computational requirements. This hierarchical approach mirrors how practitioners intuitively investigate neural networks, moving from high-level patterns to specific neuron-level mechanisms.

For the AI interpretability community, PLOT's efficiency gains directly enable larger-scale mechanistic studies. Faster localization speeds allow researchers to investigate increasingly complex models where comprehensive DAS searches become prohibitively expensive. The framework also provides actionable guidance for targeted DAS application, achieving comparable accuracy to exhaustive methods at a fraction of the computational cost.

The implications extend beyond academic interpretability research. As regulatory scrutiny of AI systems increases, efficient mechanistic understanding becomes commercially valuable. Organizations building trustworthy AI systems benefit from faster model inspection capabilities. Future work likely explores PLOT's applicability to multimodal models and real-time interpretability applications where computational efficiency directly constrains practical deployment of interpretability tools.

• →PLOT uses optimal transport coupling to localize causal variables in neural networks without exhaustive site searching, reducing computational costs substantially
• →The framework applies progressively from coarse architectural levels to fine-grained representations, enabling efficient hierarchical investigation of neural mechanisms
• →PLOT-guided DAS achieves traditional method accuracy at fraction of full DAS runtime, providing practical efficiency gains for mechanistic interpretability at scale
• →Progressive localization strategy naturally aligns with how practitioners investigate neural networks, from high-level patterns to specific neuron behaviors
• →Faster interpretability tools directly support emerging regulatory requirements and trustworthiness standards for deploying large language and vision models