Glance reranks passages from the generator's shared prefill
Original title: Glance: The Generator Is Its Own Reranker
A resident 9B backbone adds about 0.1 GiB per skill and beats bge-reranker-v2-m3 on SciFact and NFCorpus after 39 minutes of fine-tuning.
Glance adds a lightweight decision adapter and pointer head so one generator scores candidate passages from a shared prefill, avoiding a separate reranker and per-pair prefills. After 39 minutes of MS MARCO fine-tuning it reports nDCG@10 of 0.741 on SciFact and 0.345 on NFCorpus, compared with 0.721 and 0.326 for bge-reranker-v2-m3. On 300 HotpotQA questions, its selected context improves answer F1 by 0.045 under the same base generator.
Unconfirmed: Results are described as initial and cover two retrieval datasets, one QA dataset, and one backbone family.
- Short relevance probes produce one score per passage from cached context.
- Quality keeps rising through 200 candidates, or 43k tokens, from one shared prefill.
- At 200 candidates it exceeds the cross-encoder by 0.058 nDCG@10.
- The HotpotQA F1 gain has a paired bootstrap 95% CI of 0.014 to 0.076.
From the original article
1 min read at huggingface.co
Retrieval-augmented generation commonly serves separate models for reranking and answering, duplicating model residency and passage processing. We introduce Glance, which equips the generator with a lightweight decision adapter and pointer head to score candidate passages from a shared prefill.
By Adilet Metinov, Chingiz Arziev · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.