Abstract
A system and method are described for aligning generative models with specific style and voice guidelines. Some alignment methodologies may treat generated output (e.g., text, code, images, or video) as a single unit, which can make it difficult to detect specific failures in an underlying reasoning trace. The described technology can utilize an interactive feedback loop where a human reviewer provides targeted, span-level critiques on generated content and the associated reasoning trace. In response, the model can be prompted to surface the intermediate rationale for the generated selections, which the reviewer may then adjudicate. This process can produce structured training artifacts (golden units) from the adjudicated interactions. This approach may provide a more durable mechanism for aligning personality and methodology in addition to correcting the final output.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Lummus, Kate; Zucker, Nathaniel; and Collier, Ryan, "System for Generative Model Alignment Using Adjudicated Reasoning from Span-Level Critiques", Technical Disclosure Commons, (October 28, 2026)
https://www.tdcommons.org/dpubs_series/12505