Abstract
Personalization in chat assistants requires multiple interdependent components, where outputs from earlier components are consumed as inputs by downstream components. Manual optimization of these context windows remains highly complex. To address this issue, a method is disclosed wherein context windows are organized into a hierarchical schema of atomic blocks. A joint optimization algorithm is utilized to simultaneously refine top-level inputs, atomic block roles, and child block orderings. The optimization process is guided by past candidate performance, structured evaluation rubrics and processes, and internal model signals, including attention weights and introspection scores. Overall performance is maximized across all components without manual prompt engineering or a significant reliance on textual gradients.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Andreica, Mugurel-Ionut and Mathews, Rajiv, "Multi-Component Automatic Context Window Optimization for Personalization in Large Language Models", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/11456