New Method BOHM Enables Zero-Cost Attribution for Multi-Component AI Systems Without API Access

Researchers introduce BOHM, a hierarchical attribution method that extracts component importance directly from routing weights in compound AI systems, eliminating the need for expensive coalition evaluations required by Shapley-based methods like SHAP. BOHM achieves near-parity with SHAP on benchmarks (Kendall tau 0.928 vs 0.980) while requiring 9,000x fewer evaluations and working with opaque third-party APIs and agentic orchestrators where traditional methods fail.

Why it matters: As AI systems grow more complex with specialized components and external APIs, practitioners need interpretability methods that work without full system access—BOHM solves this scalability and privacy problem while providing multi-resolution insights across hierarchical component levels simultaneously.