Learn your summarization guidelines from the data instead of hand-writing prompts: +11.0 ROUGE-1 and +18.2 ROUGE-2 across five output formats

arXiv 2608.04307·medium signalXBlueskyLinkedInCopy link

MIDAS is a multi-LLM iterative framework for enterprise summarization — support tickets, legal documents, incident reports — where output must follow strict domain-specific guidelines. Rather than manual prompt engineering, it infers patterns from the data and personalizes to the use case, delivering up to 11.0% ROUGE-1, 18.2% ROUGE-2, and 8.0% ROUGE-L improvement plus consistent BERTScore F1 gains on customer-ticket summarization across five output formats, with results replicated in finance. Accepted at ICDAR 2026; the reusable move is deriving the style contract from labeled examples instead of encoding it by hand in a system prompt.