BOHM: Zero-Cost Attribution for Compound AI Systems
A new method called BOHM extracts hierarchical attribution from routing weights in compound AI systems, offering zero-cost, multi-resolution analysis without accessing component internals. This addresses limitations of Shapley-based methods like SHAP, which require evaluating arbitrary component subsets—impossible for third-party APIs and opaque endpoints. BOHM computes leaf attribution as the product of root-to-leaf routing weights and level-k attribution as the distribution over depth-k nodes. The paper, published on arXiv (2605.22866), presents BOHM as a practical alternative for deployed agentic orchestrators.
Key facts
- BOHM extracts hierarchical attribution from routing weights.
- Zero marginal cost and no access to component internals required.
- Provides multi-resolution attribution at every level simultaneously.
- Addresses limitations of Shapley-based methods like SHAP.
- Published on arXiv with ID 2605.22866.
- Method works for third-party APIs and opaque endpoints.
- Leaf attribution is path product of root-to-leaf routing weights.
- Level-k attribution is induced distribution over depth-k nodes.
Entities
Institutions
- arXiv