Glance paper has RAG generator score its own passages, skips reranker

The generator model in a retrieval-augmented pipeline can score its own candidate passages well enough to retire a dedicated reranker, two researchers argue in a new arxiv preprint.

Adilet Metinov and Chingiz Arziev call the method Glance. It bolts what the paper describes as a "lightweight decision adapter and pointer head" onto a 9B generator so the model scores passages from its cached context, "without separately prefilling each query-passage pair." A single resident backbone, the authors write, handles generation, general decisions, and reranking at a cost of "approximately 0.1 GiB of additional memory per decision skill."

The reported numbers are modest. After 39 minutes of MS MARCO fine-tuning, Glance posts 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 a 300-question HotpotQA slice, Glance-selected context lifts answer F1 by 0.045 over cross-encoder-selected context under the same base generator, with a paired bootstrap 95% CI of [0.014, 0.076]. A scaling study pushes to 200 candidates and 43k tokens scored from one shared prefill, where the paper reports exceeding the cross-encoder by 0.058 nDCG@10.

This is a single preprint on narrow retrieval benchmarks, with no production latency figures published alongside the quality numbers. The HotpotQA result rests on 300 questions, and the generator backbone is not named.