As LLM agents call more APIs, tools, and services, it becomes harder to tell which ones actually matter. In this talk, we present AgentSHAP, a black-box method for attributing an agent’s output to its tools using Shapley values. We’ll show how AgentSHAP works in practice, how it scales via Monte Carlo sampling, and how it can be used to identify redundant tools, reduce cost, and improve agent reliability in production systems.
Miriam Horovicz is an AI researcher currently working at Fiverr, with a strong focus on language models and conversational AI. She holds a B.Sc. in Computer Science and is completing a Master’s degree in Artificial Intelligence. Miriam has previous experience at National Instruments, where she led computer vision and generative AI projects. She’s also co-authored research on explainability in large language models, including the TokenSHAP method presented at NLP4Science 2024.
Keynote session: Hadas Grossmon Ella
Break
Lightning talks session
Roundtable closing
Talk by Hila Paz
Talk by Dr. Moran Mizrahi
Closing remarks
End
