Elamur Research⌕
☰
home / papers / 2026 / AURORA: Neuro-Symbolic Continual Indexing for Evolving RAG Systems

AURORA: Neuro-Symbolic Continual Indexing for Evolving RAG Systems

Paper · 2026 · Manoj Saravanan, Rohit Kumar Salla and Ramya Manasa Amancherla

Summary

Retrieval indices in RAG pipelines are usually built once and then treated as fixed, even though queries and documents keep changing. When new topics or terminology appear, recall degrades. Rebuilding the index from scratch fixes this but scales with corpus size and is too slow for frequent updates. The authors present AURORA, which frames index upkeep as few-shot continual learning. The core idea is to separate the discrete index structure from the continuous metric used to score candidates. In the dense setting, the HNSW graph is built once and its edges are never modified. Only a neural implicit codebook with residual gating is updated, using Reptile-style meta-learning on a small support set. Candidates found through the frozen graph are then rescored with the adapted metric. A contextual-bandit router chooses among dense, SPLADE, DSI and static fallback indices based on uncertainty signals. Across dense, learned sparse and generative retrieval, the framework restores up to +26.9% Recall@10 on unseen topics and adapts in milliseconds instead of seconds. This makes keeping RAG knowledge current considerably cheaper.

Key points

  • Motivating experiment: on SIFT1M, a static index reaches 99.5% recall on in-distribution queries but only 73.0% on out-of-distribution ones.
  • Adaptation speed: absorbing a new semantic cluster takes 28 ms with AURORA versus 5149 ms for full retraining, roughly 182× faster. Recall@10 on novel topics improves by up to +26.9%.
  • Router behaviour: the router uses query entropy, max-IDF and generation confidence as features, and reaches 85% routing accuracy.
    • Without enough entropy regularization (λ = 50), it collapsed to always choosing HNSW.
  • Forgetting: Experience Replay keeps base-collection performance largely intact (0.937 → 0.919) across successive 10% data injections.
  • Stated limitations:
    • The lexical drift used in evaluation is synthetic (token replacement), not natural language change.
    • The method needs a GPU; the gating forward pass costs 577 µs on CPU.
    • Real-world reward proxies are noisy.
    • Few-shot updates could be exploited for data poisoning.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksInformation Systems · 1,437 works

Cite

Cite as Manoj Saravanan, Rohit Kumar Salla and Ramya Manasa Amancherla. 2026. AURORA: Neuro-Symbolic Continual Indexing for Evolving RAG Systems. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.495

@inproceedings{saravanan2026aurora,
  title = {{AURORA: Neuro-Symbolic Continual Indexing for Evolving RAG Systems}},
  author = {Manoj Saravanan, Rohit Kumar Salla and Ramya Manasa Amancherla},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.495},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.495}
}

References

18 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yu A Malkov and Dmitry A Yashunin. 2018. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(4):824–836.
  2. Omar Besbes, Yonatan Gur, and Assaf Zeevi. 2014. Stochastic multi-armed-bandit problem with non-stationary rewards. In Advances in Neural Information Processing Systems (NeurIPS), volume 27. find
  3. Michael McCloskey and Neal J Cohen. 1989. Catastrophic interference in connectionist networks: The sequential learning problem. The Psychology of Learning and Motivation, 24:109–165. find
  4. Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Maarten de Rijke, Wei Chen, Yixing Fan, and Xueqi Cheng. 2023. Continual learning for generative retrieval over dynamic corpora. arXiv preprint arXiv:2308.14968. — [arXiv:2308.14968](https://arxiv.org/abs/2308.14968) find
  5. Sanket Mehta, Jai Gupta, Yi Tay, Mostafa Dehghani, Vinh Q Tran, Jinfeng Rao, Marc Najork, Emma Strubell, and Donald Metzler. 2022. DSI++: Updating transformer memory with new documents. arXiv preprint arXiv:2212.09744. find
  6. Thibault Formal, Benjamin Piwowarski, and Stéphane Clinchant. 2021. SPLADE: Sparse lexical and expansion model for first stage ranking. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2263–2267.
  7. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531.
  8. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation Alex Nichol, John Schulman, and Oleg Klimov. 2018. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999. find
  9. Ian Osband, Charles Blundell, Alexander Pritzel, and Benjamin Van Roy. 2016. Deep exploration via bootstrapped dqn. In Advances in Neural Information Processing Systems (NeurIPS), volume 29. find
  10. Biswajit Paria, Chih-Kuan Hsu, Shengjia Shen, and Barnabás Póczos. 2020. Minimizing flops to learn efficient sparse representations. In International Conference on Learning Representations. find
  11. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, and 1 others. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, volume 32.
  12. Mohammad Pezeshki, Sékou-Oumar Kaba, Yoshua Bengio, Aaron Courville, Doina Precup, and Guillaume Lajoie. 2021. Gradient starvation: A learning proclivity in neural networks. In Advances in Neural Information Processing Systems (NeurIPS), volume 34, pages 1256–1272. find
  13. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392.
  14. Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. 2020. Green AI. Communications of the ACM, 63(12):54–63.
  15. HuiJeong Son, Hyeongu Kang, Sunho Kim, Subeen Ho, SeongKu Kang, Dongha Lee, and Susik Yoon. 2026. CREAM: Continual retrieval on dynamic streaming corpora with adaptive soft memory. arXiv preprint arXiv:2601.02708. — [arXiv:2601.02708](https://arxiv.org/abs/2601.02708) find
  16. Yi Tay, Vinh Q Tran, Mostafa Dehghani, Dara Bahri, Ankit Gupta, Hieu Pham, and Donald Metzler. 2022. Transformer memory as a differentiable search index. In Advances in Neural Information Processing Systems, volume 35, pages 21831–21843.
  17. Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A heterogenous benchmark for zero-shot evaluation of information retrieval models. In NeurIPS Datasets and Benchmarks.
  18. David H Wolpert and William G Macready. 1997. No free lunch theorems for optimization. IEEE transactions on evolutionary computation, 1(1):67–82. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/saravanan-2026-aurora-neuro-symbolic-continual-indexing/