Preprint / Version 1

When Abstention Gates Overtrust Noisy Evidence: An Empirical Audit of Calibration and Thresholding in Lightweight RAG Evidence-Sufficiency Models

Authors

  • Kushaagra Patel Student

DOI:

https://doi.org/10.58445/csk35g23

Keywords:

Retrieval-Augmented Generation, Evidence Sufficiency , Selective Prediction, Abstention, Probability Calibration, Conformal Methods, Explainable AI

Abstract

Retrieval-augmented generation (RAG) systems retrieve passages that could overlap with the query, but lack the evidence necessary to fully address it. In this work, I examine whether a simple evidence sufficiency gate can discern fully informative contexts from one of three regimes that I construct of insufficient evidence contexts. Using 1,000 prefiltered HotpotQA questions, I construct four different contexts per question (one sufficient, three insufficient; 4,000 contexts in all) and compare dense cosine similarity, TF-IDF logistic regression, and a four-feature logistic regression. On the final 200-question set, unmodified cosine similarity produces false sufficiency rates (FSRs) of 44.5% and 19.0% for the partial evidence and lexical trap regimes. Platt scaling lowers cosine ECE from 0.1619 to 0.0267; the chosen operating point produces FSRs of 54.5% and 25.5% for those regimes. The four-feature model obtains AUROC 0.8583, with 12.5% and 8.5% FSRs, respectively, at a Y=1 threshold of 0.50. Applying the estimated Y=1 threshold over 600 precalibrated insufficient contexts yields 15.5% and 10.5% FSRs over the test set; applying only the partial evidence and lexical trap contexts produces 8.5% and 7.5%. These results are descriptive and apply to the constructed question-clustered contexts. Since there are several contexts per question that may be interrelated, and since the procedure for estimating the decision threshold at this level has not been validated, these results do not provide any distribution free finite sample guarantee. The study examines evidence sufficiency, not LLM hallucination rates downstream.

References

P. Lewis et al., “Retrieval-augmented generation for knowledge-intensive NLP tasks,” Advances in Neural Information Processing Systems, vol. 33, pp. 9459–9474, 2020. https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html

Z. Yang et al., “HotpotQA: A dataset for diverse, explainable multi-hop question answering,” in Proceedings of EMNLP, 2018, pp. 2369–2380. https://doi.org/10.18653/v1/D18-1259

O. Yoran, T. Wolfson, O. Ram, and J. Berant, “Making retrieval-augmented language models robust to irrelevant context,” in Proceedings of ICLR, 2024. https://proceedings.iclr.cc/paper_files/paper/2024/hash/8011b23e1dc3f57e1b6211ccad498919-Abstract-Conference.html

H. Joren et al., “Sufficient context: A new lens on retrieval augmented generation systems,” in Proceedings of ICLR, 2025. https://proceedings.iclr.cc/paper_files/paper/2025/hash/33dffa2e3d2ab74a783d1a8c292f66d9-Abstract-Conference.html

J. C. Platt, “Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods,” in Advances in Large Margin Classifiers, A. J. Smola, P. Bartlett, B. Schölkopf, and D. Schuurmans, Eds. Cambridge, MA, USA: MIT Press, 1999, pp. 61–74.

C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” in Proceedings of the 34th International Conference on Machine Learning, 2017, pp. 1321–1330. https://proceedings.mlr.press/v70/guo17a.html

A. N. Angelopoulos, S. Bates, E. Fisch, L. Lei, and T. Schuster, “Conformal risk control,” in Proceedings of ICLR, 2024. https://proceedings.iclr.cc/paper_files/paper/2024/hash/f3549ef9b5ff520a7e41ff3cc306ab2b-Abstract-Conference.html

V. Vovk, A. Gammerman, and G. Shafer, Algorithmic Learning in a Random World. New York, NY, USA: Springer, 2005. https://doi.org/10.1007/b106715

R. J. Tibshirani, R. F. Barber, E. Candes, and A. Ramdas, “Conformal prediction under covariate shift,” in Advances in Neural Information Processing Systems, vol. 32, 2019. https://proceedings.neurips.cc/paper_files/paper/2019/hash/8fb21ee7a2207526da55a679f0332de2-Abstract.html

C. K. Chow, “On optimum recognition error and reject tradeoff,” IEEE Transactions on Information Theory, vol. 16, no. 1, pp. 41–46, 1970. https://doi.org/10.1109/TIT.1970.1054406

Y. Geifman and R. El-Yaniv, “Selective classification for deep neural networks,” in Advances in Neural Information Processing Systems, vol. 30, 2017. https://papers.nips.cc/paper_files/paper/2017/hash/4a8423d5e91fda00bb7e46540e2b0cf1-Abstract.html

N. Reimers and I. Gurevych, “Sentence-BERT: Sentence embeddings using Siamese BERT-networks,” in Proceedings of EMNLP-IJCNLP, 2019, pp. 3982–3992. https://doi.org/10.18653/v1/D19-1410

S. Robertson and H. Zaragoza, “The probabilistic relevance framework: BM25 and beyond,” Foundations and Trends in Information Retrieval, vol. 4, nos. 1–2, pp. 1–174, 2009. https://doi.org/10.1561/1500000019

Downloads

Posted

2026-11-15