Revitalizing Black-Box Interpretability: Actionable Interpretability for LLMs via Proxy Models

Researchers propose a cost-effective proxy model framework that uses smaller, efficient models to approximate the interpretability explanations of expensive Large Language Models (LLMs), achieving over 90% fidelity at just 11% of computational cost. The framework includes verification mechanisms and demonstrates practical applications in prompt compression and data cleaning, making interpretability tools viable for real-world LLM development.

The computational barrier to understanding how LLMs make decisions has long constrained the practical application of interpretability research. Existing model-agnostic explanation techniques require extensive computational resources that render them impractical for production systems, forcing organizations to choose between interpretability and efficiency. This research directly addresses that tension by proposing a proxy model approach where smaller, cheaper models learn to approximate the decision boundaries of expensive LLMs, dramatically reducing the cost of generating reliable explanations.

The innovation represents a shift in how the AI research community approaches the interpretability-scalability tradeoff. Rather than developing more efficient explanation algorithms, this work leverages model approximation with a statistical verification step (the screen-and-apply mechanism) to ensure proxy explanations align with actual model behavior before deployment. The empirical validation demonstrating 90%+ fidelity is significant because it suggests proxy explanations can serve as reliable guides for optimization, not merely informative visualizations.

For AI developers and organizations building with LLMs, this work transforms interpretability from a debugging curiosity into a practical development primitive. The demonstrated applications—prompt compression reducing computational overhead and automated poisoned example removal improving data quality—show concrete pathways to ROI from interpretability investments. The open-source release amplifies impact by enabling rapid adoption across research and production environments.

The framework's effectiveness hinges on the assumption that smaller models can faithfully approximate larger ones' decision boundaries. Future research should explore how this approach scales to frontier models and whether proxy explanations remain valid as target models evolve, particularly in domains where model behavior differs substantially from training distributions.

• →Proxy models achieve 90% fidelity in explaining LLM decisions while using only 11% of computational resources, making interpretability economically viable at scale.
• →Statistical verification mechanisms validate proxy explanations locally before deployment, reducing risks from model approximation errors.
• →Actionable interpretability enables practical LLM optimization including prompt compression and automated removal of poisoned training examples.
• →Open-source release of code and datasets accelerates adoption of cost-effective interpretability techniques across research and production teams.
• →Framework shifts interpretability from passive observation tool to scalable primitive that directly improves model development workflows.