Glance:生成器就是自己的重排序器
Glance: The Generator Is Its Own Reranker
Glance 將生成與重排序整合至同一個 9B backbone,透過共享預填上下文及輕量適配器和 pointer head 為候選段落評分。經 39 分鐘 MS MARCO 微調後,Glance 在 SciFact 和 NFCorpus 的 nDCG@10 分別為 0.741 和 0.345,高於 bge-reranker-v2-m3 的 0.721 和 0.326。
arxiv:2610.40454
Glance: The Generator Is Its Own Reranker
Published on Oct 29
·Submitted by
Metinov Adileton Oct 30
Authors:
Adilet Metinov,
Chingiz Arziev
Abstract
Glance integrates reranking and generation into a single backbone using lightweight adapters and pointer heads to score passages from shared context, reducing memory overhead while improving retrieval and answer quality.
Generated by thinkingmachines/Inkling-Small
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. Short relevance probes branch from the cached context, producing one score per passage without separately prefilling each query-passage pair. A single resident 9B backbone supports generation, general decisions and reranking, with approximately 0.1 GiB of additional memory per decision skill. After 39 minutes of MS MARCO fine-tuning, Glance achieves 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 over cross-encoder-selected context under the same base generator (paired bootstrap 95% CI [0.014, 0.076]). In a scaling study, reranking quality keeps rising up to 200 candidates (43k tokens) scored from one shared prefill, where it exceeds the cross-encoder by 0.058 nDCG@10, and a probe-only activated adapter lets generation continue from the original passage cache. These initial results demonstrate generator-based reranking with one resident backbone; evaluation currently covers two retrieval datasets, one QA dataset and one backbone family.
View arXiv pageView PDFProject pageAdd to collection
Community
![]()
Paper submitter44 minutes ago
Glance: one model. Judge, rank, generate.
RAG usually serves two models: a cross-encoder reranker that encodes every (query, passage) pair, then a generator that reads the winners again. We show the generator can be its own reranker. All k passages are prefilled once into a shared state, one short probe per passage continues from that cache, and a LoRA + pointer head scores every passage in the same pass. With the adapter off, the same weights chat as usual.
On a 9B hybrid (Gated DeltaNet) backbone:
- 39 min of training on one B200: SciFact nDCG@10 0.741, NFCorpus 0.345 (bge-reranker-v2-m3: 0.721 / 0.326)
- Scales to k=200 candidates: 0.717 vs 0.659 for BGE (SciFact subset)
- HotpotQA answer F1 +0.045 over BGE-selected context, same generator (paired bootstrap CI excludes 0)
- Activated variant: chat stays bit-identical to the base model, and the answer continues the reranking cache
Adapters + code (Apache-2.0): https://huggingface.co/airun-ai/GlanceRAG-MiMo-9B-Reranker
Reply
Edit Preview
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Comment ·Sign up or log in to comment
Models citing this paper 3
#### airun-ai/MiMo-V2.6-Distill-Qwen-9B-Decision Updated about 1 hour ago
#### airun-ai/GlanceRAG-MiMo-9B-Reranker Updated about 1 hour ago
#### airun-ai/GlanceRAG-MiMo-9B-Reranker-Activated Updated about 1 hour ago
Datasets citing this paper 0
No dataset linking this paper
Cite arxiv.org/abs/2610.40454 in a dataset README.md to link it from this page.
Spaces citing this paper 0
No Space linking this paper
Cite arxiv.org/abs/2610.40454 in a Space README.md to link it from this page.
Collections including this paper 0
No Collection including this paper
Add this paper to a collection to link it from this page.
System theme
Company
Website
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co
