BOHM: Zero-Cost Hierarchical Attribution for Compound AI Systems

In three linesBOHM is a hierarchical attribution method for compound AI systems that extracts component contributions directly from routing weights without evaluating arbitrary subsets. Zero marginal cost, compatible with opaque third-party APIs. On 18 LLMs (880 LiveCodeBench problems), Kendall tau=0.928 vs SHAP tau=0.980 at 9,000x more evaluations.

Summary generated by Claude — human-verified