Ticket-trained model improves computing support answers, but agent orchestration falls short

A preprint reports that training a language model on rewritten help-desk resolutions improved its answers to high-performance computing support questions, while combining it with documentation search failed to improve on the individual tools. Researchers trained Llama-3-8B using 11,244 historical query-resolution pairs and evaluated it on 400 held-out tickets from Purdue's Rosen Center for Advanced Computing. Agreement with the center's resolutions, measured by BERTScore F1, rose from 0.06 to 0.27 over the base model. Models run on institutional infrastructure because ticket data may not leave the institution. The combined assistant followed the documentation tool in 78% of answers; neither a learned selector nor a language-model verifier beat the trained tool alone.