airun-ai/GlanceRAG-MiMo-9B-Reranker-Activated

Glance-Activated — MiMo-9B Reranker

One model. Judge, rank, generate.

One model, one read of the context: rerank every passage, then answer from the same cache.

This is the activated Glance reranker for XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B. The LoRA adapter is off while the model reads the retrieved passages (the shared state) and on only for the short relevance probes, one per passage, that continue from the state's cache, in the spirit of Activated LoRA (Greenewald et al., 2025). Consequences:

1. Chat is exactly the base model. Ordinary text never reaches the adapter, so with this adapter loaded the model generates bit-for-bit like the base.
2. The reranking cache is the base model's cache. After the probes score every passage, the answer is generated by continuing the same cache: the passages are read once for both reranking and answering.

Rank-16 LoRA + 256-d pointer head, warm-started from airun-ai/MiMo-V2.6-Distill-Qwen-9B-Decision, trained 150 steps (1,200 MS MARCO records, 17 minutes on one H100) with the activated patch. Not a standalone model; requires the patch in code/alora_patch.py (call apply() before the first forward).

Paper: Glance: The Generator Is Its Own Reranker (Adilet Metinov, Chingiz Arziev; Airun Intelligence Labs, 2026). arXiv: 2610.40454.

Results

nDCG@10, BM25 top-20, packed_all, one pass:

Adapter	SciFact	NFCorpus
BM25	0.662	0.306
bge-reranker-v2-m3	0.721	0.326
Control: LoRA on all tokens, same 150 steps	0.744	0.351
Activated (this adapter)	0.740	0.351

Cache reuse on 100 HotpotQA distractor questions (10 paragraphs, median state 1.4k tokens, H100): state pass 64 ms, 10 probes 85 ms (recall@3 0.97). Answering by continuing the reranking cache: EM 0.61 / F1 0.750, answer latency 138 ms; re-prefilling the top-3 passages instead: EM 0.59 / F1 0.754, 181 ms.

Correctness check (Qwen3.5-0.8B, randomised LoRA): with the patch, the adapter-on state cache equals the base model's bitwise across all 48 cache tensors (attention KV, DeltaNet conv and recurrent states).

Code

• code/alora_patch.py: the activated-adapter patch for the Kev runtime (training and inference).
• code/train_alora.py: training entry (wraps kev.train, requires --shared_prefix 1).
• code/check_alora.py: bitwise cache-identity check.
• code/chat_alora.py: chat-identity check vs the base.
• code/e1_rerank_alora.py: BEIR reranking evaluation (--alora 1).
• code/e5_cache_reuse.py: HotpotQA rerank-then-continue-the-cache evaluation.

Runtime: Kev at commit 0a6d3faf; see requirements.txt.

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.