Abstract
Reinforcement learning for fine-tuning generative models can be negatively impacted by noisy reward signals from single-score automated evaluation systems, which may lead to training instability and unpredictable model improvements. A framework for dynamic reward synthesis addresses this by decomposing high-level concepts into discrete, objective attributes. By evaluating these attributes individually and weighting them based on the context of the input, the system produces a reward signal that is more stable, interpretable, and aligned with specific goals.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Weisz, Agoston; Goldenberg, Roman; and Sluzhaev, Evgeny, "Dynamic Reward Synthesis via Contextually Weighted Objective Attribute Raters", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/9741