Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 25·8 min read
BOHM: Zero-Cost Hierarchical Attribution for Compound AI Systems
TL;DR
Compound AI systems (multiple specialized components working together) need attribution methods to explain which components contributed to outputs, but standard Shapley-based methods like SHAP require testing all possible component combinations—impossible with third-party APIs or opaque systems. BOHM extracts attribution directly from routing weights already maintained by these systems at zero additional cost.
✦ Why It Matters
Engineers can now attribute compound AI system outputs without expensive recomputation or API access, enabling real-time explainability for production agentic systems.
Key Takeaways
- Use BOHM for efficient attribution in complex AI systems.
- Leverage existing routing weights for multi-resolution insights.
- Achieve high correlation with ground-truth rankings at low cost.
- Consider BOHM as a complementary tool alongside SHAP.
How It Works
BOHM constructs a hierarchical attribution tree by utilizing the routing weights that AI systems maintain. It calculates leaf attribution as the product of routing weights from the root to each leaf node, while level-k attribution reflects the distribution of contributions across nodes at a specific depth.
This method allows for simultaneous analysis at multiple levels without needing to evaluate all possible component combinations.