Elamur Research⌕
☰
home / papers / 2026 / Quantifying and Improving the Robustness of Retrieval-Augmented Language Models Against Spurious Features in Grounding Data

Quantifying and Improving the Robustness of Retrieval-Augmented Language Models Against Spurious Features in Grounding Data

Paper · 2026 · Shiping Yang, Jie Wu, Wenbiao Ding et al.

Summary

Retrieval-augmented language models (RALMs) are usually tested against noise that changes a document's meaning. This paper examines a quieter problem: changes that leave the content intact but alter surface properties, such as writing style, origin, chunk order, markup format or metadata, can still flip a model's answer. The authors call these spurious features. They introduce SURE, an automated perturb-then-evaluate framework built on NQ-Open queries, a Wikipedia dump and the Contriever-msmarco retriever. SURE injects spurious features with model-based and rule-based edits. It then confirms the meaning is unchanged using bidirectional entailment and string matching, and scores models instance by instance. The hardest cases form a compact benchmark, SIG_Wiki. Experiments show the sensitivity is common across LLMs, and chain-of-thought-style prompting does not fix it. Fine-tuning on SURE-generated data does fix much of it, which makes the framework useful both for diagnosis and for training more reliable RAG readers.

Key points

  • Taxonomy: five spurious-feature types split into 13 perturbations:
    • Style: Simple, Complex
    • Source: LLM-generated, Self-generated
    • Logic: Random, LLM Rank, Reverse
    • Format: HTML, YAML, Markdown, JSON
    • Meta: Source, Time
  • Metrics: Win Rate (wrong→right), Lose Rate (right→wrong) and Robustness Rate (answer unchanged).
  • Prompting methods fall short: Chain-of-Note scored below plain Qwen2.5-72B-Instruct on several categories. DeepSeek-R1 was less robust than DeepSeek-V3 across all five types.
  • Attention analysis: in cases where the answer changed, attention to the answer span in the document shifted significantly more than in robust cases (Welch's t-test, p = 0.046).
  • Training fixes: SFT and DPO used over 30k synthetic samples, two epochs, on Llama-3.1-8B-Instruct.
    • On SIG_Wiki, most categories rose from 10–24% robustness to about 94–99.5%.
    • SFT did best in-domain, while DPO generalized better to the out-of-domain SIG_Trivial benchmark.
    • Code is available at github.com/maybenotime/RAG-SpuriousFeatures.

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

 Institutions

🏢MicrosoftRedmond · United States378 works · 161,630 citations🎓University of Hong KongPok Fu Lam · China118 works · 13,496 citations🔬Canadian Institute for Advanced ResearchToronto · Canada35 works · 19,316 citations

Also on the byline: Atlassian.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 worksInformation Systems · 1,730 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksInformation retrieval & search · 552 works

Cite

Cite as Shiping Yang, Jie Wu, Wenbiao Ding, Ning Wu, Shining Liang, Ming Gong, Hongzhi Li, Hengyuan Zhang, Angel X. Chang and Dongmei Zhang. 2026. Quantifying and Improving the Robustness of Retrieval-Augmented Language Models Against Spurious Features in Grounding Data. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1545

@inproceedings{yang2026quantifying,
  title = {{Quantifying and Improving the Robustness of Retrieval-Augmented Language Models Against Spurious Features in Grounding Data}},
  author = {Shiping Yang, Jie Wu, Wenbiao Ding, Ning Wu, Shining Liang, Ming Gong, Hongzhi Li, Hengyuan Zhang, Angel X. Chang and Dongmei Zhang},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1545},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1545}
}

References

49 references, 25 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, and 1 others. 2016. Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268.
  2. Ning Bian, Hongyu Lin, Peilin Liu, Yaojie Lu, Chunkang Zhang, Ben He, Xianpei Han, and Le Sun. 2024. Influence of external information on large language models mirrors social cognitive patterns. IEEE Transactions on Computational Social Systems. find
  3. Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. 2024a. Benchmarking large language models in retrieval-augmented generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17754–17762.
  4. Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023b. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997.
  5. Xiaoyang Chen, Ben He, Hongyu Lin, Xianpei Han, Tianshu Wang, Boxi Cao, Le Sun, and Yingfei Sun. 2024b. Spiral of silences: How is large language model killing information retrieval?–a case study on open domain question answering. arXiv preprint arXiv:2404.10496. — [arXiv:2404.10496](https://arxiv.org/abs/2404.10496) find
  6. Yunfan Gao, Yun Xiong, Yijie Zhong, Yuxi Bi, Ming Xue, and Haofen Wang. 2025. Synergizing rag and reasoning: A systematic review. arXiv preprint arXiv:2504.15909. — [arXiv:2504.15909](https://arxiv.org/abs/2504.15909) find
  7. Sukmin Cho, Soyeong Jeong, Jeongyeon Seo, Taeho Hwang, and Jong C. Park. 2024. Typos that broke the RAG’s back: Genetic attack on RAG pipeline by simulating documents in the wild via low-level perturbations. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 2826–2844, M… find
  8. Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. 2024. The power of noise: Redefining retrieval for rag systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Developm…
  9. Sunhao Dai, Yuqi Zhou, Liang Pang, Weihao Liu, Xiaolin Hu, Yong Liu, Xiao Zhang, Gang Wang, and Jun Xu. 2024. Neural retrievers are biased towards llm-generated content. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 526–537. find
  10. Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The faiss library. arXiv preprint arXiv:2401.08281.
  11. Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. 2024. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625–630. find
  12. Philip Feldman, James R Foulds, and Shimei Pan. 2024. Ragged edges: The double-edged sword of retrieval-augmented chatbots. arXiv preprint arXiv:2403.01193. — [arXiv:2403.01193](https://arxiv.org/abs/2403.01193) find
  13. Chunjing Gan, Dan Yang, Binbin Hu, Hanxiao Zhang, Siyuan Li, Ziqi Liu, Yue Shen, Lin Ju, Zhiqiang Zhang, Jinjie Gu, and 1 others. 2024. Similarity is not all you need: Endowing retrieval augmented generation with multi layered thoughts. arXiv preprint arXiv:2405.19893. — [arXiv:2405.19893](https://arxiv.org/abs/2405.19893) find
  14. Mingqi Gao, Jie Ruan, Renliang Sun, Xunjian Yin, Shiping Yang, and Xiaojun Wan. 2023a. Human-like summarization evaluation with chatgpt. arXiv preprint arXiv:2304.02554.
  15. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
  16. Jia He, Mukund Rungta, David Koleczek, Arshdeep Sekhon, Franklin X Wang, and Sadid Hasan. 2024. Does prompt formatting have any impact on llm performance? arXiv preprint arXiv:2411.10541. — [arXiv:2411.10541](https://arxiv.org/abs/2411.10541) find
  17. Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2021. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.09118.
  18. Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2023. Atlas: Few-shot learning with retrieval augmented language models. Journal of Machine Learning Research, 24(251):1–43.
  19. Yuxuan Jiang, Dawei Li, and Frank Ferraro. 2025. Drp: Distilled reasoning pruning with skill-aware step decomposition for efficient large reasoning models. arXiv preprint arXiv:2505.13975.
  20. Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601–1611.
  21. Aobo Kong, Shiwan Zhao, Hao Chen, Qicheng Li, Yong Qin, Ruiqi Sun, Xin Zhou, Enzhi Wang, and Xiaohang Dong. 2024. Better zero-shot reasoning with role-play prompting. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language…
  22. Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations.
  23. Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent retrieval for weakly supervised open domain question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6086–6096.
  24. In The Twelfth International Conference on Learning Representations. find
  25. Cheng Li, Jindong Wang, Kaijie Zhu, Yixuan Zhang, Wenxin Hou, Jianxun Lian, and Xing Xie. 2023. Emotionprompt: Leveraging psychology for large language models enhancement via emotional stimulus. arXiv e-prints, pages arXiv–2307. find
  26. Jiejun Tan, Zhicheng Dou, Wen Wang, Mang Wang, Weipeng Chen, and Ji-Rong Wen. 2024a. Htmlrag: Html is better than plain text for modeling retrieved knowledge in rag systems. arXiv preprint arXiv:2411.02959. — [arXiv:2411.02959](https://arxiv.org/abs/2411.02959) find
  27. Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, and 1 others. 2025. From generation to judgment: Opportunities and challenges of llm-as-a-judge. In Proceedings of the 2025 Conference on Empirical Methods… find
  28. Zhen Tan, Dawei Li, Song Wang, Alimohammad Beigi, Bohan Jiang, Amrita Bhattacharjee, Mansooreh Karami, Jundong Li, Lu Cheng, and Huan Liu. 2024b. Large language models for data annotation and synthesis: A survey. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processi…
  29. Dawei Li, Shu Yang, Zhen Tan, Jae Young Baik, Sunkwon Yun, Joseph Lee, Aaron Chacko, Bojian Hou, Duy Duong-Tran, Ying Ding, and 1 others. 2024. Dalk: Dynamic co-augmentation of llms and kg to answer alzheimer’s disease questions with scientific literature. arXiv preprint arXiv:2405.04819. find
  30. Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157–173.
  31. Weihao Liu, Ning Wu, Shiping Yang, Wenbiao Ding, Shining Liang, Ming Gong, and Dongmei Zhang. 2025. Mudaf: Long-context multi-document attention focusing through contrastive learning on attention heads. arXiv preprint arXiv:2502.13963. — [arXiv:2502.13963](https://arxiv.org/abs/2502.13963) find
  32. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837.
  33. Jinyang Wu, Feihu Che, Chuyuan Zhang, Jianhua Tao, Shuai Zhang, and Pengpeng Shao. 2024a. Pandora’s box or aladdin’s lamp: A comprehensive analysis revealing the role of rag noise in large language models. arXiv preprint arXiv:2408.13533.
  34. Kevin Wu, Eric Wu, and James Zou. 2024b. Clasheval: Quantifying the tug-of-war between an llm’s internal prior and external evidence. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track.
  35. Yi Liu, Lianzhe Huang, Shicheng Li, Sishuo Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. 2023. Recall: A benchmark for llms robustness against external counterfactual knowledge. arXiv preprint arXiv:2311.08147. — [arXiv:2311.08147](https://arxiv.org/abs/2311.08147) find
  36. Peng Xu, Wei Ping, Xianchao Wu, Lawrence McAfee, Chen Zhu, Zihan Liu, Sandeep Subramanian, Evelina Bakhturina, Mohammad Shoeybi, and Bryan Catanzaro. 2024. Retrieval meets long context large language models. In The Twelfth International Conference on Learning Representations.
  37. Yannic Neuhaus, Maximilian Augustin, Valentyn Bor-eiko, and Matthias Hein. 2023. Spurious features everywhere-large-scale detection of harmful spurious features in imagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20235–20246. find
  38. Shiping Yang, Renliang Sun, and Xiaojun Wan. 2023a. A new benchmark and reverse validation method for passage-level hallucination detection. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3898–3908. find
  39. Pouya Pezeshkpour and Estevam Hruschka. 2024. Large language models sensitivity to the order of options in multiple-choice questions. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 2006–2017.
  40. Shiping Yang, Renliang Sun, and Xiaojun Wan. 2023b. A new dataset and empirical study for sentence simplification in chinese. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8306– 8321. find
  41. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728–53741.
  42. Melanie Sclar, Yejin Choi, Yulia Tsvetkov, and Alane Suhr. 2024. Quantifying language models’ sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting. find
  43. Kui Yu, Xianjie Guo, Lin Liu, Jiuyong Li, Hao Wang, Zhaolong Ling, and Xindong Wu. 2020. Causality-based feature selection: Methods and evaluations. ACM Computing Surveys (CSUR), 53(5):1–36. find
  44. Wenhao Yu, Hongming Zhang, Xiaoman Pan, Kaixin Ma, Hongwei Wang, and Dong Yu. 2023. Chain-of-note: Enhancing robustness in retrieval-augmented language models. arXiv preprint arXiv:2311.09210.
  45. Hengyuan Zhang, Chenming Shang, Sizhe Wang, Dongdong Zhang, Yiyao Yu, Feng Yao, Renliang Sun, Yujiu Yang, and Furu Wei. 2025. Shifcon: Enhancing non-dominant language capabilities with a shift-based multilingual contrastive framework. In Proceedings of the 63rd Annual Meeting of the Association for…
  46. Hengyuan Zhang, Zhihao Zhang, Mingyang Wang, Zun-hai Su, Yiwei Wang, Qianli Wang, Shuzhou Yuan, Ercong Nie, Xufeng Duan, Qibo Xue, and 1 others. 2026. Locate, steer, and improve: A practical survey of actionable mechanistic interpretability in large language models. arXiv preprint arXiv:2601.14004.
  47. Yujia Zhou, Yan Liu, Xiaoxi Li, Jiajie Jin, Hongjin Qian, Zheng Liu, Chaozhuo Li, Zhicheng Dou, Tsung-Yi Ho, and Philip S Yu. 2024. Trustworthiness in retrieval-augmented generation systems: A survey. arXiv preprint arXiv:2409.10102. — [arXiv:2409.10102](https://arxiv.org/abs/2409.10102) find
  48. Kaijie Zhu, Jindong Wang, Jiaheng Zhou, Zichen Wang, Hao Chen, Yidong Wang, Linyi Yang, Wei Ye, Yue Zhang, Neil Zhenqiang Gong, and 1 others. 2023. Promptbench: Towards evaluating the robustness of large language models on adversarial prompts. arXiv e-prints, pages arXiv–2306. find
  49. Jingming Zhuo, Songyang Zhang, Xinyu Fang, Haodong Duan, Dahua Lin, and Kai Chen. 2024. Prosa: Assessing and understanding the prompt sensitivity of llms. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 1950– 1976. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yang-2026-quantifying-improving-robustness-retrieval-augmented/