Researchers fine-tune 406M model for meeting summaries with retrieved text spans
A new paper on arXiv introduces a method called Retrieved-Span Training to improve query-focused meeting summarization. The approach fine-tunes a 406M-parameter Fusion-in-Decoder model on retrieved text spans of up to 2,000 words instead of full transcripts. When tested on the QMSum dataset, the model achieves a 36.33 ROUGE-1 score, compared to 35.41 for a 1.2B-parameter baseline. The authors…
Key points
- 406M-parameter model fine-tuned on 2,000-word retrieved text spans scores 36.33 ROUGE-1 on QMSum
- Smaller model uses one-third fewer parameters and half the memory of a 1.2B baseline
- Replacing first 4,500 transcript words with 2,000 retrieved words boosts performance by 1.55 ROUGE-1
The paper also shows that replacing the first 4,500 words of a transcript with 2,000 retrieved words improves performance by 1.55 ROUGE-1 on test data. Under a single prompt and scorer, the 406M model outperforms five proprietary hosted models by at least 6.2 ROUGE-1, though the results lack human or factuality evaluation. The findings are limited to QMSum and automated metrics.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Study finds wiki-indexed LLM outperforms vector RAG for cross-unit questions in ML courses · 1 src
- Study proposes skill habits to fix AI agent inconsistency · 1 src
- Researchers propose latent equivalence learning for enterprise data agents, score 94.67% on benchmark · 1 src
- Researchers extract circuits from language models using Attention routing · 1 src
- ReAdapt improves warm‑introduction and reaction selection accuracy for Gemini‑3‑Flash · 1 src
Comments
via GitHub Discussions