Abstract
Generating long-form documents from large, diverse data sources can present challenges for some generative systems, which may exhibit unstable optimization and difficulty isolating various error types, such as factual hallucination or poor evidence selection. A disclosed technology can address these potential issues using a decoupled, two-stage framework. An offline stage can perform multi-objective optimization by distilling source data and a candidate draft into intermediate representations for selection salience, informational semantics, style, and factual provenance, which can provide granular feedback for refining generative prompts. In an online stage, these optimized prompts can be compiled for efficient, low-latency document generation. This separation of computationally intensive analysis from production inference may improve the factual grounding, stylistic coherence, and overall quality of the generated documents by facilitating more stable, dimension-specific optimization.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Badjatiya, Pinkesh, "Decoupled Multi-Objective Optimization for Document Generation Using Intermediate Representations", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12072