Abstract
Conventional supervised fine-tuning for specialized image generation models can be inefficient, potentially relying on manual, intuition-driven experimentation that may slow iteration and have difficulty systematically incorporating detailed quality feedback. An automated system can use a large language model (LLM) agent that may operate in a closed-loop optimization process. The agent can receive a high-level goal and structured metadata describing available datasets and tuning parameters to generate an executable training configuration. The system may also interpret detailed, categorical quality assurance reports that identify specific failure modes, for example, "shape change." By reasoning about the causes of these failures, the LLM can propose targeted configuration adjustments for subsequent iterations, facilitating a systematic, data-driven approach to discover improved model configurations more efficiently.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Cunningham, Jacob; Yang, Alan; Kemelmacher-Shlizerman, Ira; and Lee, Chris, "LLM Agent to Interpret Categorical Quality Feedback", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/9940