Results (BM25 top-20, nDCG@10)
Table with columns: Method, SciFact (300 q), NFCorpus (323 q)| Method | SciFact (300 q) | NFCorpus (323 q) |
|---|
| BM25 | 0.662 | 0.306 |
| bge-reranker-v2-m3 | 0.721 | 0.326 |
| Decision 9B zero-shot, 20 separate calls | 0.729 | 0.323 |
| Glance zero-shot, one pass | 0.715 | 0.341 |
Glance trained (checkpoints/msmarco-short), one pass | 0.741 | 0.345 |
| Glance trained-long (this adapter), one pass | 0.749 | 0.346 |
Scaling (SciFact, 100 queries, 100-word passages): at k=100 candidates (21.6k tokens) in one pass, this adapter reaches nDCG@10 0.697 vs 0.637 for bge-reranker-v2-m3 on the same candidates (1.05 s per query on a B200), and at k=200 (43k tokens, chunked probe batches) 0.717 vs 0.659.
HotpotQA distractor, 300 questions (checkpoints/msmarco-short), same 9B generator answering from the top-3: context selected by this adapter gives F1 0.729 vs 0.684 for bge-reranker-v2-m3 (adapter off while answering; +0.045, 95% CI [+0.014, +0.076]) and recalls more gold paragraphs (R@2 0.897 vs 0.863).
One resident backbone serves base chat, the general decision adapter and this reranking adapter, switched per request in about 10 ms, with outputs bit-identical to each loaded alone (16.85 GiB total).
Use
python -m pip install -r requirements.txtpython -m pip install --no-deps triton==3.8.0
Rerank and answer with one backbone (code/glance_rag.py):
from huggingface_hub import snapshot_downloadfrom code.glance_rag import Glancerag = Glance(snapshot_download("airun-ai/GlanceRAG-MiMo-9B-Reranker")) # local dir of this reposcores = rag.rerank(query, passages) # one packed forward pass, adapter onanswer = rag.answer(query, passages, top_n=3) # same weights, adapter off
An activated variant (adapter only on probe tokens, exact base-model chat, answer generated from the reranking cache) is at airun-ai/GlanceRAG-MiMo-9B-Reranker-Activated. A vLLM prototype serving chat and decisions in one engine is in code/vllm_proto/.
Evaluation scripts: code/e1_rerank.py (BEIR reranking modes), code/e4_hotpot.py (HotpotQA end to end), code/multi_mode.py (three modes on one backbone), code/toggle_test.py (real-time on/off switching), code/make_train.py (training records), code/e5_scaling.py (k scaling, batching), code/e6_order.py (order sensitivity). Training uses the Kev runtime (python -m kev.train, see training_config.json).
Limitations
Proof of concept: small evaluation sets, one backbone family. Latency is from an unoptimised eager path; a cross-encoder is faster per query (24–60 ms vs ~270–420 ms on a B200). Trained on MS MARCO; review the MS MARCO terms of use for your application. Memory currently grows with k × context length (56 GiB at k=100); k=200 does not fit in one call and runs on Kev's chunked path (15.8 s per query).
Citation
@misc{metinov2026glance, title = {Glance: The Generator Is Its Own Reranker}, author = {Metinov, Adilet and Arziev, Chingiz}, year = {2026}, eprint = {2610.40454}, archivePrefix = {arXiv}, primaryClass = {cs.CL}, url = {https://arxiv.org/abs/2610.40454}}
License: Apache-2.0. See LICENSE and NOTICE; redistributions must retain the NOTICE attribution.