Revolutionizing AI Attribution: Introducing BOHM

Published on May 25, 2026

Attribution methods in machine learning have relied heavily on Shapley-based approaches, which assess contributions of different components in hierarchical systems. These techniques typically require evaluating the system on various subsets, posing challenges for complex setups involving third-party APIs and opaque endpoints. As AI systems continue to grow in complexity, a new method was needed.

Researchers have unveiled BOHM, an innovative attribution method that sidesteps the limitations of traditional Shapley methods. BOHM leverages existing routing weights to build a hierarchical attribution tree without needing access to internal components. This approach not only simplifies the process but also provides multi-resolution attribution simultaneously, a feature previously unattainable with flat methods.

Testing BOHM against 18 large language models demonstrated its effectiveness, yielding a Kendall tau of 0.928 compared to Shapley’s 0.980, albeit with significantly fewer evaluations—9,000 times less. In scenarios where routing decisions are less straightforward, the method outperformed expectations, showing that drivers concentrate on a limited selection of tools while maintaining robust performance across several benchmarks.

BOHM’s introduction brings crucial advancements to AI systems, enabling efficient and meaningful attribution analysis without excessive resource demands. of efficiency and providing actionable insights on routing decisions, it stands to enhance the understanding of AI behavior in real-world applications. Furthermore, it offers a complementary approach to existing methods, highlighting the evolving landscape of AI transparency and accountability.

Related News