This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Optimas: Optimizing Compound AI Systems with Globally Aligned Local Rewards
Created by
Haebom
Author
Shirley Wu, Parth Sarthi, Shiyu Zhao, Aaron Lee, Herumb Shandilya, Adrian Mladenic Grobelnik, Nurendra Choudhary, Eddie Huang, Karthik Subbian, Linjun Zhang, Diyi Yang, James Zou, Jure Leskovec
Outline
In this paper, we propose the Optimas framework to solve the optimization problem of complex AI systems that integrate multiple components (e.g., large-scale language models, specialized tools, and traditional machine learning models). Optimas maintains a Local Reward Function (LRF) for each component, and is designed such that each LRF is correlated with the global system performance (local-global alignment). At each iteration, Optimas efficiently adapts and maximizes the LRF of each component while maintaining this correlation. This enables independent updates of heterogeneous components, and ensures that local improvements lead to global performance improvements. Through extensive evaluations on five real-world complex systems, we demonstrate an average performance improvement of 11.92% over existing methods.
Takeaways, Limitations
•
Takeaways:
◦
We present a general and effective framework for efficient optimization of complex AI systems.
◦
Enables independent optimization of heterogeneous configurations of components (prompts, hyperparameters, model parameters, etc.).
◦
It showed significant performance improvement over existing methods in real complex systems.
•
Limitations:
◦
A detailed description of the design of LRF and maintenance of local-global alignment may be lacking.
◦
Further research may be needed to determine generalizability to different types of complex systems.
◦
An analysis of the computational cost and complexity of the adaptation process of LRF may be required.
