Efficient AI Attribution via Existing Model Weights
2 min read
Certainly, modern AI often uses many small parts working together. However, understanding how each part contributes is hard. Therefore, it usually requires running the system thousands of times, which is slow and expensive.
In particular, BOHM offers a new way. Essentially, it uses the routing weights the system already has to find answers. Thus, it provides multi-resolution insights at zero-cost, without needing access to the internal parts. Hence, it is a faster, complementary tool.
For example, tests show BOHM gives results very close to complex methods but much faster. Importantly, it works even when some parts are private or unknown. Consequently, this method helps developers and researchers see how their compound AI systems make decisions efficiently.
| Aspect | BOHM (Proposed) | SHAP (Shapley-Based) |
|---|---|---|
| Cost & Requirements | Zero marginal cost; requires only routing weights, no access to component internals. | High cost; requires evaluating the system on many arbitrary component subsets. |
| Attribution Type | Hierarchical, multi-resolution (provides attribution at every level of the component tree simultaneously). | Flat; provides a single vector of per-component marginal contributions. |
| Scalability for Complex Systems | Highly scalable; computable wherever routing state exists, even for large hierarchies (e.g., 475 leaves). | Poor scalability; often infeasible for large systems, third-party APIs, or agentic orchestrators. |
| Key Convergence | Agrees with SHAP when the deployed router is near-optimal. | Serves as the optimal benchmark; its agreement with BOHM signals routing quality. |
| Best Use Case | A complementary primitive for quick, diagnostic, multi-level understanding of compound AI systems. | The gold-standard for precise, axiom-based feature attribution when full evaluation is possible. |
BOHM: Zero-Cost Hierarchical Attribution
In addition, BOHM offers zero-cost hierarchical attribution for compound AI systems. Consequently, it uses the systems’ routing weights to assign importance without extra computation. As a result, everyone can apply it even with opaque APIs. Therefore, it provides multi-resolution insights that flat methods cannot. Moreover, it complements existing techniques like SHAP. Furthermore, its efficiency makes it accessible to diverse developers.
Enabling Attribution in Opaque Systems
“BOHM satisfies efficiency, monotonicity, symmetry, and weak suppression but not Shapley’s additivity. It is best understood as a complementary primitive: a multi-resolution decomposition computable wherever routing state exists, whose disagreement with Shapley is itself diagnostic.”
Ultimately, BOHM offers a practical, zero-cost way to understand compound AI systems. In conclusion, its hierarchical approach lets everyone—from researchers to everyday users—see how decisions are made. Looking ahead, this method can make AI attribution more accessible and transparent for all communities.
Ultimately, BOHM provides a new way to understand how compound AI systems work by looking at their internal routing. Consequently, this method assigns credit to each part of the system without needing extra computations. Therefore, it is especially useful for complex systems where traditional methods are too costly or impossible.
In conclusion, the approach shows strong alignment with established methods on standard tests. Thus, it offers developers a practical tool for system diagnosis and improvement. Accordingly, BOHM enables inclusive analysis of AI pipelines that use various external tools and services.
