Abstract
A system for managing modular directives for large language models can address challenges associated with redundant or conflicting instructions. The system can create a latent policy manifold by vectorizing a repository of natural language instructions and applying a matrix factorization technique, such as singular value decomposition, to identify an underlying orthogonal basis. This manifold can be used by an offline governance component to analyze and reduce redundancy in the instruction repository and by an online routing engine to select a subset of instructions at runtime. The selected subset of instructions can be both relevant to a given query and mathematically orthogonal. This method may improve model efficiency, reduce context window saturation, and contribute to more predictable behavior in generative artificial intelligence systems that rely on dynamic composition of behavioral directives.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kim, Nathan Myung-Chul and Addisie, Abraham, "Latent Policy Manifold for Governing and Selecting Orthogonal AI Directives", Technical Disclosure Commons, (July 07, 2026)
https://www.tdcommons.org/dpubs_series/10822