- What changed
- A 406M Fusion-in-Decoder model fine-tuned on 2,000-word retrieved spans recovers 6.30 ROUGE-1 loss from capped input and matches a 1.2B system on QMSum with lower inference memory.
- Why you should care
- Shows compact models can match larger ones in meeting summarization with span-based input tuning.
- Your move
- Test. Results are QMSum-specific and based on automatic metrics only.
- What to watch next
- Human evaluation of factuality and coherence in retrieved-span meeting summaries on QMSum or similar datasets.
- Event
- research
- Event date
- Sep 23, 2026
- Relevant to
- General AI readers
