Elamur Research⌕
☰
home / papers / 2026 / Rational Synthesizers or Heuristic Followers? Analyzing LLMs in RAG-based Question-Answering

Rational Synthesizers or Heuristic Followers? Analyzing LLMs in RAG-based Question-Answering

Paper · 2026 · Atharv Naphade

Summary

The paper asks how LLMs in retrieval-augmented generation (RAG) settle on an answer when the retrieved context holds many documents that disagree with each other. Earlier benchmarks such as ConflictingQA only tested one-versus-one document conflicts. The author built GroupQA to cover the group case. GPT-4o generated contested Yes/No questions across 95 topic categories. Supporting and opposing web documents were collected through the Google Search API, and two LLM judges labeled each document's stance.

With this resource, the study manipulates groups of evidence by varying quantity, redundancy, ordering and the presence of conflict. It measures answer flip thresholds, belief plasticity and leave-one-out influence for open and closed models. The models turn out to follow shallow cues rather than weigh evidence rationally:

  • Repeating one paraphrased argument sways them more than independent sources do.
  • Documents placed early carry extra weight.
  • Bigger models adapt less to context.
  • Chain-of-thought explanations do not faithfully reflect what drove the answer.

These results point to manipulation risks such as context stuffing. They also suggest defenses: de-duplicating evidence, shuffling document order and deliberately retrieving dissenting views.

Key points

  • GroupQA dataset: 1,635 controversial binary questions with 15,058 evidence documents, averaging about 9.2 per question and roughly 384 words per paragraph. 1,635 of 1,883 processed questions were accepted (86.83%), drawn from 22,264 scraped documents.
  • Collection and labeling pipeline:
    • Each question was turned into affirmative and negative assertions, and the top 10 Google results were retrieved for each.
    • Text was extracted with Trafilatura, and metadata was removed.
    • Stance was assigned only when GPT-4o-mini and Gemini-2.5-Flash agreed.
    • Strength labels (Strong/Medium/Weak) were annotated but not used in the experiments.
    • Human checks on a sample showed about 99% agreement.
  • Illusory Truth Effect: a single paraphrased document repeated several times flips answers more than distinct supporting documents, for example with Llama-3 70B.
  • Primacy and conflict effects:
    • Evidence presented first has outsized influence on the answer.
    • Balanced, explicitly conflicting contexts stabilize decisions and reduce the advantage of redundancy, though they do not remove it.
  • Scope limits: English-only web text, binary questions, no source metadata, and no mechanistic analysis of model internals. Data and code are available at github.com/anaphade/GroupQA.

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

🎓Carnegie Mellon UniversityPittsburgh · United States539 works · 179,533 citations

 Fields, topics and concepts

📚 Shelves 6The 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 worksComputer Vision and Pattern Recognition · 2,869 worksInformation Systems · 1,437 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksQuestion answering · 441 works

Cite

Cite as Atharv Naphade. 2026. Rational Synthesizers or Heuristic Followers? Analyzing LLMs in RAG-based Question-Answering. 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.2003

@inproceedings{naphade2026rational,
  title = {{Rational Synthesizers or Heuristic Followers? Analyzing LLMs in RAG-based Question-Answering}},
  author = {Atharv Naphade},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.2003},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.2003}
}

References

23 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet 40301 9 find
  2. Deshpande. 2024. Geo: Generative engine optimization. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5–16. ArXiv:2311.09735. — [arXiv:2311.09735](https://arxiv.org/abs/2311.09735) find
  3. Shakiba Amirshahi and 1 others. 2024. Evaluating the robustness of retrieval-augmented generation to adversarial evidence in the health domain. arXiv preprint arXiv:2509.03787. — [arXiv:2509.03787](https://arxiv.org/abs/2509.03787) find
  4. Anthony Chen and Wen-tau Yih. 2022. Rich knowledge sources for open-domain question answering. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics. find
  5. Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, and Ruifeng Xu. 2024. Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),…
  6. Linfeng Gao and 1 others. 2024. Probing latent knowledge conflict for faithful retrieval-augmented generation. arXiv preprint arXiv:2510.12460. — [arXiv:2510.12460](https://arxiv.org/abs/2510.12460) find
  7. Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: Retrieval-augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning.
  8. Lynn Hasher, David Goldstein, and Thomas Toppino. 1977. Frequency and the conference of referential validity. Journal of Verbal Learning and Verbal Behavior, 16(1):107–112. find
  9. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems, volum…
  10. Han Liu, Yupeng Zhang, Bingning Wang, Weipeng Chen, and Xiaolin Hu. 2024a. Full-ECE: A metric for token-level calibration on large language models. arXiv preprint arXiv:2406.11345. — [arXiv:2406.11345](https://arxiv.org/abs/2406.11345) find
  11. Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024b. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12:157–173.
  12. Shayne Longpre, Kartik Perisetla, Anthony Chen, Chris Ramesh, Nikhil andlue, and John Schulman. 2021. Entity-based knowledge conflicts in question answering. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7052–7063.
  13. Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048–11064.
  14. Mrinank Sharma, Meg Tong, Tomasz Korbak, David Rogers, N Bennett, Amanda andjb, and 1 others. 2023. Understanding and mitigating sycophancy in large language models. arXiv preprint arXiv:2310.13548.
  15. Alexander Wan, Eric Wallace, and Dan Klein. 2024. What evidence do language models find convincing? In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7468–7483, Bangkok, Thailand. Association for Computational Linguistics. find
  16. Jinzhu Wang, Zhen Xu, Di Jin, Xin Yang, and Tao Li. 2025. Accommodate knowledge conflicts in retrieval-augmented llms: Towards reliable response generation in the wild. In Proceedings of AAAI Conference on Artificial Intelligence. find
  17. Allan Wei, Jerry andr Dafoe and Jason Wei. 2023. Simple synthetic data reduces sycophancy in large language models. arXiv preprint arXiv:2308.03958. find
  18. Chong Xiang and 1 others. 2024. Certifiably robust rag against retrieval corruption. arXiv preprint arXiv:2405.15556. find
  19. Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. 2023. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453. find
  20. Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. 2024. Efficient streaming language models with attention sinks. In Proceedings of the 12th International Conference on Learning Representations. Key aliased to match text. find
  21. Rongwu Xu, Zehan Shi, Zhuo Wang, Ningyu Yan, Feiliang Zhu, Yunsong Yao, and Xiaoyan Li. 2024. Knowledge conflicts for llms: A survey. arXiv preprint arXiv:2403.08319.
  22. Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. 2024. Making retrieval-augmented language models robust to irrelevant context. arXiv preprint arXiv:2310.01558. Published at ICLR 2024.
  23. Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, and 1 others. 2023. Siren’s song in the ai ocean: A survey on hallucination in large language models. arXiv preprint arXiv:2309.01219. 40302 10

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/naphade-2026-rational-synthesizers-or-heuristic-followers/