Glance: The Generator Is Its Own Reranker
One 9B model ranks passages and generates answers, adding about 0.1 GiB of memory per decision skill.
Glance scores candidate passages through short relevance probes branching from a shared cached context, avoiding separate processing of each query-passage pair. After 39 minutes of fine-tuning, it beat bge-reranker-v2-m3 on SciFact and NFCorpus. On 300 HotpotQA questions, its selected context improved answer F1 by 0.045 over cross-encoder-selected context using the same generator. The results remain limited to two retrieval datasets, one question-answering dataset and one backbone family.
HF Daily Papers' note
Glance scores candidate passages through short relevance probes branching from a shared cached context, avoiding separate processing of each query-passage pair. After 39 minutes of fine-tuning, it beat bge-reranker-v2-m3 on SciFact and NFCorpus. On 300 HotpotQA questions, its selected context improved answer F1 by 0.045 over cross-encoder-selected context using the same generator. The results remain limited to two retrieval datasets, one question-answering dataset and one backbone family.
HF Daily Papers' note
score 5