📄 GenAIus at RAG4Reports 2026: Citation-Aware Compression for Multilingual Report Generation

Summary

The paper presents the GenAIus team's entry to the RAG4Reports 2026 Multilingual Report Generation Task. In this task, English report requests must be answered with reports in which every sentence is backed by citations to a news corpus in English, Chinese, Russian and Arabic. The topics overlapped with those of TREC 2025 RAGTIME, so the authors kept their earlier evidence pipeline fixed and reused its outputs: question decomposition, multilingual retrieval with cross-question score aggregation, LLM-extracted nuggets, and LLM-based thematic clustering. They changed only the writing stage. Instead of producing the short and long reports separately, the system first writes the long report from the clustered nuggets. It then edits that report down to the short version, keeping the most important claims and only the citations that still support them. All runs relied on prompting alone, with no fine-tuning. With GPT-4o, this compression approach placed second on the leaderboard and improved both citation support and coverage over the team's independent-generation baseline. The result suggests that length-constrained grounded reporting benefits from separating synthesis from content selection.

Key points

• The corpus was sampled from Common Crawl News (2021–2024) in four languages. Evaluation used the Auto-ARGUE framework, which scores how well reports cover nuggets and how well their sentences are backed by citations, along with an F1 that combines the two.
• The baseline run, trial (GPT-4o, reports generated independently, as in the earlier genaius-cluster TREC run), ranked 3rd with F1 0.5382, sentence support 0.7936 and nugget coverage 0.4395.
• The best run, genaius-cluster-gpt4 (GPT-4o, long report then citation-aware compression), ranked 2nd with F1 0.5456, support 0.8140 and coverage 0.4428.
• The genaius-cluster-gpt5 run swapped in GPT-5.4. It reached the highest coverage (0.4487), but its support fell to 0.6620, giving F1 0.5006 and 5th place. This illustrates the tradeoff between coverage and citation faithfulness.
• Future work will focus on stronger citation verification and on systematic ablations of the generation stage.