Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

This paper claims that an energy-based scoring head over dense embeddings is a more reliable abstention signal than softmax probability for RAG systems in women's health, especially on near-domain confusable queries.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

An energy-based scoring head trained on dense embeddings improves abstention decisions for medical RAG systems on semantically hard out-of-distribution queries compared to softmax and kNN baselines.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A modest but well-run study: an energy scoring head edges out softmax on synthetic hard OOD abstention in a women's-health RAG, but the same-generator test set makes the real-world claim contingent. the 4 major comments →

arxiv 2509.04482 v2 pith:CERNBCAM submitted 2025-08-31 cs.CL cs.AI

Energy Landscapes Enable Reliable Abstention in Retrieval-Augmented Large Language Models for Healthcare

Ravi Shankar , Sheng Wong , Lin Li , Magdalena Bachmann , Alex Silverthorne , Beth Albert , Gabriel Davis Jones This is my paper
classification cs.CL cs.AI
keywords energy-based modelsabstentionretrieval-augmented generationout-of-distribution detectionsafe AI in healthcarewomen's healthnegative samplinghallucination prevention
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a small energy-based scoring head, trained over frozen dense text embeddings, can decide when a retrieval-augmented question-answering system in women's health should refuse to answer, and that this energy signal is more reliable than the usual softmax probability on the most dangerous kind of mistake: queries that look medically plausible but fall outside the indexed guideline corpus. On a test set of such confusable near-distribution questions, the energy model reports AUROC 0.961 versus 0.950 for a calibrated softmax baseline and drops the false-positive rate at 95% recall from 0.331 to 0.235. The authors show that this robustness survives ablations in which negative sampling changes, as long as both easy and hard negatives are present during training, and that removing the energy head collapses the advantage. If true, this gives RAG systems a cheap pre-generation filter that can defer hard queries to retrieval expansion or a human, without the cost of generating multiple candidate answers.

Core claim

The paper's central claim is that a scalar energy score—one small learned head on top of a 256-dimensional projection of frozen dense embeddings—separates in-scope clinical questions from out-of-scope lookalikes more reliably than the probability output of a softmax classifier trained on the same data. On the hard split of their test set, where each in-domain question is paired with a synthetic near-distribution paraphrase that swaps specialty terms, the energy model reports AUROC 0.961 and false-positive rate at 95% recall 0.235, versus 0.950 and 0.331 for softmax; the detection error drops about 30% relative. The paper further claims this robustness comes from the energy head itself, not f

What carries the argument

The central mechanism is the energy-calibrated semi-contrastive triplet loss (EC-SCTL), which trains the shared projector and the energy head jointly. For each anchor question it uses a positive, a deterministically paired hard negative, and a small set of uniformly sampled easy/mid-range negatives; the loss sums a cosine-similarity margin and an energy margin per negative, then aggregates across negatives with a temperature-controlled LogSumExp. The energy head is a two-layer feedforward network that outputs a scalar; low energy marks in-domain queries. The same projector also feeds a softmax classification head for the baseline, so the comparison isolates the scoring rule.

Load-bearing premise

The load-bearing premise is that synthetic hard negatives, built by swapping obstetric terms for analogous terms from other specialties and then filtered by a medical language model, faithfully represent the real near-domain out-of-scope questions that a deployed RAG system will face.

What would settle it

Take a set of real out-of-scope obstetric questions written or collected by clinicians (e.g., paediatric gynaecology or oncology questions), embed them, and measure EBM versus softmax AUROC; if the advantage disappears or reverses on these real queries, the synthetic benchmark was not representative. A cheaper check: vary the filtering threshold used to select hard negatives after generation and confirm the reported margin persists.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A RAG system in women's health can run this energy model before generation and defer any query whose energy exceeds a threshold calibrated on validation data.
  • Because the energy score is a single forward pass over a small network, abstention adds negligible latency compared to post-hoc methods that decode multiple answers.
  • The ablation results imply that training data must include both easy and hard negatives, so practitioners cannot omit confusable examples even if they are expensive to collect.
  • The reported degradation of the kNN density baseline on hard cases suggests that retrieval systems should not use raw retrieval score as a proxy for safety.
  • The EBM advantage on near-domain confusables, if it holds, motivates separating 'easy' and 'hard' out-of-distribution splits in all future abstention benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the same energy head is applied to the retrieved passages rather than only the query, it may also fire when evidence is sparse or contradictory, giving a second, independent abstention trigger—the paper does not test this.
  • The approach should transfer to other safety-critical verticals with curated, versioned corpora, but the hard-negative generation procedure (domain-term substitution plus medical-model filtering) must be rebuilt per specialty, so the reported numbers are specific to women's health.
  • A direct test of the benchmark's validity is to evaluate on real clinician-written out-of-scope questions; the authors note synthetic negatives may not capture real-world complexity.
  • Since the representations are frozen, retraining the underlying embedding encoder on the clinical corpus could change the energy landscape's margin; the paper leaves this end-to-end question open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes an energy-based abstention (EBM) mechanism for retrieval-augmented generation (RAG) in women's health. A 256-dimensional projector maps BGE-M3 embeddings of clinical questions into a shared space, with an energy head and a softmax classifier trained under a semi-contrastive loss that combines cosine-similarity and energy margins. Training uses reciprocal-nearest-neighbour positives, synthetically generated hard negatives (term substitution, e.g., uterus→prostate), mid-range negatives, and external OOD examples from MedMCQA/SQuAD. The EBM is compared with a softmax head and a kNN density baseline on easy, hard, and mixed abstention splits, with ablations over negative types. The central claim is that EBM achieves superior abstention on hard, near-distribution confusable queries (AUROC 0.961 vs. 0.950 softmax; lower FPR@95), while performing comparably on easy OOD.

Significance. If the reported results hold, the work offers a practical, pre-generation abstention signal that is cheaper than semantic entropy and more robust than softmax confidence on confusable queries. The design choices—energy-calibrated triplet loss, controlled negative exposure, explicit separation of easy/hard OOD evaluation, and comparison under fair data exposure—are sensible and potentially useful for safety-critical RAG. The paper also honestly lists limitations. However, the empirical support for the central claim is weakened by an internal numerical inconsistency, an ablation claim contradicted by the paper's own table, missing reported ablations, the absence of error bars, and the fact that the hard-OOD test set is generated by the same synthetic procedure as the training negatives. These issues must be addressed before the central claim can be accepted.

major comments (4)
  1. [Abstract; Table 1] The abstract reports FPR@95 of 0.235 (EBM) vs. 0.331 (softmax) on hard cases, but Table 1 (All Negs) reports 0.265 vs. 0.358. These are the headline safety-relevant numbers, and the discrepancy is not cosmetic. Please correct the abstract or the table and state the correct values.
  2. [§2.2, §2.6, §4 Discussion] The hard-OOD test set is constructed with the same prompt-based term-substitution and MedGemma filtering procedure used to create training hard negatives. The test is therefore a held-out sample from the training generator, not an independent sample of real-world near-distribution queries. The Discussion acknowledges that synthetic hard negatives 'may not capture the full complexity of real-world queries.' This undermines the external-validity claim that the method 'reduces over-commitment on queries that are most likely to elicit persuasive but unsafe responses.' Please evaluate on an independent hard/OOD set (e.g., clinician-written out-of-scope questions or other specialty guidelines) or substantially temper the generalization claim.
  3. [Table 1; §3 Results] All numbers in Table 1 are point estimates with no error bars, confidence intervals, or number of seeds. The key hard-OOD AUROC difference is 0.012 (0.961 vs. 0.949/0.950), and the FPR@95 difference comes from a single evaluation. Without multiple runs or bootstrap intervals, the reported advantage over softmax could be within noise. Please report variance across seeds or resampling.
  4. [Abstract; §2.6; Table 1] The abstract claims that 'the inclusion or exclusion of specific negative types (hard, easy, mixed) sharpens decision boundaries but is not essential for generalisation to hard cases.' Table 1 contradicts this: training on Easy Only gives hard-OOD AUROC 0.512 (EBM) and No Hard gives 0.507, versus 0.961 with all negatives. Hard negatives are clearly essential. Additionally, §2.6 states that ablations 'no energy' and 'no ext ood' were conducted, but no such rows appear in Table 1 or in the Results text; the claim that robustness stems primarily from the energy scoring head is therefore not supported by the reported data. Please report these ablations or revise the claims.
minor comments (5)
  1. [Abstract] Typo: 'superior abstention performance abstention' should be 'superior abstention performance'.
  2. [Figures] Figure 1 and Figure 2 are referenced with captions but the figures themselves are not visible in the manuscript. Please include the actual plots.
  3. [§3 Results] The sentence 'AUROC dropped from 0.970 (all negs) to 0.787 (easy only) and 0.803 (no easy)' mixes numbers from different models/columns without specifying method; clarify which rows are being compared and define 'No Hard'/'No Easy' in the table.
  4. [§2.3] The softmax head is described as a 'linear classifier (256→256→2) with GELU activation'; a GELU-activated layer is not linear. Please clarify the architecture.
  5. [Abstract; §2.1] The abstract refers to a 'dense semantic corpus of 2.6M guideline-derived questions,' while §2.1 describes selecting a representative set of 100K anchors. Clarify the relationship between the 2.6M pool and the 100K anchors.

Circularity Check

0 steps flagged

No circularity found: the EBM vs softmax comparison is a held-out empirical evaluation, and the self-citations are motivational rather than load-bearing.

full rationale

The paper's central claim is empirical: an energy head trained on dense embeddings gives better abstention AUROC on semantically hard OOD queries than a softmax head trained under the same exposure budget. The evaluation is a standard held-out split: hard negatives are synthetically generated and MedGemma-filtered, then split across training, validation, and test. The hard-OOD AUROC is computed on test pairs not used in training, and the thresholds are calibrated on validation and fixed before test (§2.6). AUROC itself is threshold-free, so no fitted threshold is renamed as the reported metric. The same generative procedure for train and test negatives raises an external-validity question—whether the synthetic 'hard' distribution represents real clinical near-distribution queries—but this is not circularity: the model is not evaluated on its own training labels, and the test is a genuine within-distribution generalization test. The paper itself flags this limitation: 'Synthetic hard negatives, though clinically plausible, may not capture the full complexity of real-world queries' (Discussion). This weakens the real-world inference but does not make the internal comparison circular. The self-citations [7,8] are used to motivate the problem (ChatGPT overconfidence in women's health, semantic entropy as a prior uncertainty signal) and are not used to justify the EBM's design or to supply the benchmark numbers. No uniqueness theorem, ansatz, or fitted parameter is imported from those works. Accordingly, no step in the derivation reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

7 free parameters · 3 axioms · 0 invented entities

No new physical or conceptual entities are introduced. The energy score is a learned scalar output of a neural network, not a new conserved quantity or dimension. The main unstated dependencies are the quality of the BGE-M3 embeddings, the representativeness of the synthetic hard-negative distribution, and the unverified hyperparameter settings.

free parameters (7)
  • T (softplus and LogSumExp temperature) = not reported
    Appears in all softplus margins and in the LogSumExp aggregation; value is not given in the manuscript text.
  • m_sim (cosine similarity margin) = not reported
    Margin in Lsim, appears in the similarity triplet term.
  • m_E (energy margin) = not reported
    Margin in Lenergy, controls energy separation between positives and negatives.
  • lambda (energy term weight) = not reported
    Scales the energy term in Lenergy.
  • w_OOD (external OOD hinge weight) = not reported
    Weight on the auxiliary OOD hinge loss.
  • w_HN (paired hard-negative hinge weight) = not reported
    Weight on the paired hard-negative hinge loss.
  • k_mine (number of sampled negatives per anchor) = not reported
    Controls the exposure budget of additional negatives per anchor; exact value not stated.
axioms (3)
  • domain assumption BGE-M3 embeddings provide a suitable semantic representation space for clinical question similarity.
    Section 2.1 embeds all texts with BAAI/bge-m3 and L2-normalizes; the entire method operates on these precomputed embeddings, so the method inherits the quality of that embedding space.
  • ad hoc to paper Synthetic hard negatives generated by term substitution are representative of real near-distribution out-of-scope queries.
    Section 2.2 defines hard negatives by substituting O&G terms with analogues from other specialties; Section 2.6 evaluates on these negatives. The paper itself notes in Limitations that synthetic negatives may not capture real-world complexity.
  • domain assumption Reciprocal-nearest-neighbour filtering yields semantically aligned positive pairs.
    Section 2.2 retains anchor-positive pairs only when each is the other's top-1 neighbour under cosine similarity, assuming this strict reciprocity indicates true semantic alignment and removes duplicates.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Energy Landscapes Enable Reliable Abstention in Retrieval-Augmented Large Language Models for Healthcare." pith.science (2026). https://pith.science/paper/CERNBCAM

@misc{pith2026250904482,
  author       = {Pith},
  title        = {Pith review of: Energy Landscapes Enable Reliable Abstention in Retrieval-Augmented Large Language Models for Healthcare},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CERNBCAM}},
  note         = {Machine review of arXiv:2509.04482}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Reliable abstention is critical for retrieval-augmented generation (RAG) systems, particularly in safety-critical domains such as women's health, where incorrect answers can lead to harm. We present an energy-based model (EBM) that learns a smooth energy landscape over a dense semantic corpus of 2.6M guideline-derived questions, enabling the system to decide when to generate or abstain. We benchmark the EBM against a calibrated softmax baseline and a k-nearest neighbour (kNN) density heuristic across both easy and hard abstention splits, where hard cases are semantically challenging near-distribution queries. The EBM achieves superior abstention performance abstention on semantically hard cases, reaching AUROC 0.961 versus 0.950 for softmax, while also reducing FPR@95 (0.235 vs 0.331). On easy negatives, performance is comparable across methods, but the EBM's advantage becomes most pronounced in safety-critical hard distributions. A comprehensive ablation with controlled negative sampling and fair data exposure shows that robustness stems primarily from the energy scoring head, while the inclusion or exclusion of specific negative types (hard, easy, mixed) sharpens decision boundaries but is not essential for generalisation to hard cases. These results demonstrate that energy-based abstention scoring offers a more reliable confidence signal than probability-based softmax confidence, providing a scalable and interpretable foundation for safe RAG systems.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Remembering More, Risking More: Longitudinal Safety Risks in Memory-Equipped LLM Agents

    cs.AI 2026-05 unverdicted novelty 6.0

    Memory-equipped LLM agents exhibit increasing safety violation rates as memory accumulates across independent tasks, termed temporal memory contamination, detected via a new trigger-probe protocol.

Reference graph

Works this paper leans on

37 extracted references · 15 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: Advances in Neural Information Processing Systems (NeurIPS), pp

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., K¨ uttler, H., Lewis, M., Yih, W.-t., Rockt¨ aschel, T., Riedel, S., Kiela, D.: Retrieval- augmented generation for knowledge-intensive nlp tasks. In: Advances in Neural Information Processing Systems (NeurIPS), pp. 9459–9474 (2020). https://proceedings.neurips.cc/paper/2020/file/6b4...

    work page 2020

  2. [2]

    Nature Medicine29, 1930– 1940 (2023)

    Thirunavukarasu, A.J., Ting, D.S.J., Elangovan, K., Gutierrez, L., Tan, T.F., Ting, D.S.W.: Large language models in medicine. Nature Medicine29, 1930– 1940 (2023)

    work page 1930

  3. [3]

    arXiv preprint arXiv:2303.13375 (2023)

    Nori, H., King, N., McKinney, S.M., Carignan, D., Horvitz, E.: Capabilities of gpt-4 on medical challenge problems. arXiv preprint arXiv:2303.13375 (2023)

    Pith/arXiv arXiv 2023

  4. [4]

    IEEE Transactions on Information Theory16(1), 41–46 (1970) https://doi.org/10.1109/TIT.1970

    Chow, C.K.: On optimum recognition error and reject tradeoff. IEEE Transactions on Information Theory16(1), 41–46 (1970) https://doi.org/10.1109/TIT.1970. 1054406

    doi:10.1109/tit.1970 1970

  5. [5]

    In: Advances in Neural Information Processing Systems, vol

    Geifman, Y., El-Yaniv, R.: Selective classification for deep neural net- works. In: Advances in Neural Information Processing Systems, vol. 30, 12 pp. 4878–4887 (2017).https://papers.nips.cc/paper/7073-selective-classification- for-deep-neural-networks

    work page 2017

  6. [6]

    In: Advances in Neural Information Processing Systems (NeurIPS), vol

    Liu, W., Wang, X., Owens, J.D., Li, Y.: Energy-based out-of- distribution detection. In: Advances in Neural Information Processing Systems (NeurIPS), vol. 33 (2020). arXiv preprint arXiv:2010.03759. https://papers.neurips.cc/paper/2020/file/f5496252609c43eb8a3d147ab9b9c006- Paper.pdf

    Pith/arXiv arXiv 2020

  7. [7]

    npj Women’s Health2(1), 26 (2024)

    Bachmann, M., Duta, I., Mazey, E., Cooke, W., Vatish, M., Davis Jones, G.: Exploring the capabilities of chatgpt in women’s health: obstetrics and gynaecology. npj Women’s Health2(1), 26 (2024)

    work page 2024

  8. [8]

    Penny-Dimri, J.C., Bachmann, M., Cooke, W.R., Mathewlynn, S., Dockree, S., Tolladay, J., Kossen, J., Li, L., Gal, Y., Jones, G.D.: Reducing large lan- guage model safety risks in women’s health using semantic entropy (2025) arXiv:2503.00269 [cs.LG]

    arXiv 2025

  9. [9]

    El-Yaniv, R., Wiener, Y.: On the Foundations of Noise-free Selective Classification vol. 11, pp. 1605–1641 (2010).https://jmlr.org/papers/v11/el-yaniv10a.html

    work page 2010

  10. [10]

    In: Proceedings of the 36th International Conference on Machine Learning, pp

    Geifman, Y., El-Yaniv, R.: Selectivenet: A deep neural network with an integrated reject option. In: Proceedings of the 36th International Conference on Machine Learning, pp. 2151–2159 (2019). https://doi.org/10.48550/arXiv.1901.09192 . https://proceedings.mlr.press/v97/geifman19a.html

    Pith/arXiv arXiv doi:10.48550/arxiv.1901.09192 2019

  11. [11]

    In: Proceedings of the 34th International Conference on Machine Learning, pp

    Guo, C., Pleiss, G., Sun, Y., Weinberger, K.Q.: On calibration of modern neu- ral networks. In: Proceedings of the 34th International Conference on Machine Learning, pp. 1321–1330 (2017). https://doi.org/10.48550/arXiv.1706.04599 . https://proceedings.mlr.press/v70/guo17a.html

    Pith/arXiv arXiv doi:10.48550/arxiv.1706.04599 2017

  12. [12]

    arXiv preprint arXiv:2107.07511 (2021) https://doi.org/10.48550/arXiv.2107.07511

    Angelopoulos, A.N., Bates, S.: A gentle introduction to conformal prediction and distribution-free uncertainty quantification. arXiv preprint arXiv:2107.07511 (2021) https://doi.org/10.48550/arXiv.2107.07511

    Pith/arXiv arXiv doi:10.48550/arxiv.2107.07511 2021

  13. [13]

    In: International Conference on Learning Representations (2017)

    Hendrycks, D., Gimpel, K.: A baseline for detecting misclassified and out- of-distribution examples in neural networks. In: International Conference on Learning Representations (2017). https://doi.org/10.48550/arXiv.1610.02136 . https://arxiv.org/abs/1610.02136

    Pith/arXiv arXiv doi:10.48550/arxiv.1610.02136 2017

  14. [14]

    In: International Conference on Learning Representations (ICLR) (2018).https://openreview.net/forum?id=H1VGkIxRZ

    Liang, S., Li, Y., Srikant, R.: Enhancing the reliability of out-of-distribution image detection in neural networks. In: International Conference on Learning Representations (ICLR) (2018).https://openreview.net/forum?id=H1VGkIxRZ

    work page 2018

  15. [15]

    In: Advances in Neural Information Processing Systems (NeurIPS), pp

    Lee, K., Lee, K., Lee, H., Shin, J.: A simple unified framework for detect- ing out-of-distribution samples and adversarial attacks. In: Advances in Neural Information Processing Systems (NeurIPS), pp. 7167–7177 (2018). 13 https://arxiv.org/pdf/1807.03888

    Pith/arXiv arXiv 2018

  16. [16]

    Mohseni, S., Pitale, M., Yadawa, J., Wang, Z.: Self-supervised learning for generalizable out-of-distribution detection34(4), 5216–5223 (2020)

    work page 2020

  17. [17]

    Yang, J., Zhou, K., Li, Y., Liu, Z.: Generalized out-of-distribution detection: A survey (2024) arXiv:2110.11334 [cs.CV]

    Pith/arXiv arXiv 2024

  18. [18]

    Predicting Structured Data1, 1–59 (2006)

    LeCun, Y., Chopra, S., Hadsell, R., Ranzato, M., Huang, F.J.: A tutorial on energy-based learning. Predicting Structured Data1, 1–59 (2006)

    work page 2006

  19. [19]

    In: International Conference on Learning Representations (2020)

    Grathwohl, W., Wang, K.-C., Jacobsen, J.-H., Duvenaud, D., Norouzi, M., Swersky, K.: Your classifier is secretly an energy based model and you should treat it like one. In: International Conference on Learning Representations (2020). https://doi.org/10.48550/arXiv.1912.03263 . https://openreview.net/forum?id=Hkxzx0NtDB

    Pith/arXiv arXiv doi:10.48550/arxiv.1912.03263 2020

  20. [20]

    In: Proceedings of the 58th Annual Meeting of the Association for Computa- tional Linguistics (ACL), pp

    Kamath, A., Jia, R., Liang, P.: Selective question answering under domain shift. In: Proceedings of the 58th Annual Meeting of the Association for Computa- tional Linguistics (ACL), pp. 5684–5696 (2020). https://doi.org/10.18653/v1/ 2020.acl-main.503 .https://aclanthology.org/2020.acl-main.503/

    doi:10.18653/v1/ 2020

  21. [21]

    Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., Schiefer, N., Hatfield-Dodds, Z., DasSarma, N., Tran-Johnson, E., Johnston, S., et al.: Language models (mostly) know what they know (2022) arXiv:2207.05221 [cs.CL]

    Pith/arXiv arXiv 2022

  22. [22]

    In: The Twelfth International Conference on Learning Representations (2024)

    Asai, A., Wu, Z., Wang, Y., Sil, A., Hajishirzi, H.: Self-rag: Learning to retrieve, generate, and critique through self-reflection. In: The Twelfth International Conference on Learning Representations (2024). https://doi.org/10.48550/arXiv. 2310.11511 .https://arxiv.org/abs/2310.11511

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.11511 2024

  23. [23]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Schroff, F., Kalenichenko, D., Philbin, J.: Facenet: A unified embedding for face recognition and clustering. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 815–823 (2015). https://doi.org/ 10.1109/CVPR.2015.7298682

    arXiv 2015

  24. [24]

    (2018).https://arxiv.org/abs/1706.07567

    Wu, C.-Y., Manmatha, R., Smola, A.J., Kr¨ ahenb¨ uhl, P.: Sampling matters in deep embedding learning. (2018).https://arxiv.org/abs/1706.07567

    Pith/arXiv arXiv 2018

  25. [25]

    (2017).https://arxiv.org/abs/1703.07737

    Hermans, A., Beyer, L., Leibe, B.: In defense of the triplet loss for person re- identification. (2017).https://arxiv.org/abs/1703.07737

    Pith/arXiv arXiv 2017

  26. [26]

    In: Pro- ceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp

    Karpukhin, V., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., Yih, W.-t.: Dense passage retrieval for open-domain question answering. In: Pro- ceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 6769–6781 (2020). https://doi.org/10.18653/v1/2020. emnlp-main.550 .https://aclanthology.org/2020.emnlp-...

    doi:10.18653/v1/2020 2020

  27. [27]

    In: International Conference on Learning Representations (2021).https://openreview.net/forum?id=zeFrfgyZln

    Xiong, L., Xiong, C., Li, Y., Tang, K.-F., Liu, J., Bennett, P.N., Ahmed, J., Overwijk, A.: Approximate nearest neighbor negative contrastive learning for dense text retrieval. In: International Conference on Learning Representations (2021).https://openreview.net/forum?id=zeFrfgyZln

    work page 2021

  28. [28]

    In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp

    Rajpurkar, P., Jia, R., Liang, P.: Know what you don’t know: Unanswerable ques- tions for squad. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp. 784–789 (2018). https://aclanthology.org/P18-2124/

    work page 2018

  29. [29]

    Artificial Intelligence in Medicine102, 101753 (2020)

    Zhang, H., Zhang, Y., Wang, Y., Chen, J., Li, X., Liang, Y.: A survey on deep learning for medical question answering. Artificial Intelligence in Medicine102, 101753 (2020)

    work page 2020

  30. [30]

    In: Proceed- ings of the Conference on Health, Inference, and Learning (CHIL), pp

    Pal, A., Umapathi, L.K., Sankarasubbu, M.: Medmcqa: A large-scale multi- subject multi-choice dataset for medical domain question answering. In: Proceed- ings of the Conference on Health, Inference, and Learning (CHIL), pp. 248–260. PMLR, ??? (2022).https://proceedings.mlr.press/v174/pal22a.html

    work page 2022

  31. [31]

    In: Proceedings of the 2016 Conference on Empir- ical Methods in Natural Language Processing (EMNLP), pp

    Rajpurkar, P., Zhang, J., Lopyrev, K., Liang, P.: Squad: 100,000+ questions for machine comprehension of text. In: Proceedings of the 2016 Conference on Empir- ical Methods in Natural Language Processing (EMNLP), pp. 2383–2392 (2016). https://aclanthology.org/D16-1264/

    work page 2016

  32. [32]

    (2022).https://arxiv.org/abs/2204.06507

    Sun, Y., Ming, Y., Zhu, X., Li, Y.: Out-of-distribution detection with deep nearest neighbors. (2022).https://arxiv.org/abs/2204.06507

    Pith/arXiv arXiv 2022

  33. [33]

    In: Advances in Neural Information Processing Sys- tems, vol

    Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., Raffel, C.A.: Mixmatch: A holistic approach to semi-supervised learning. In: Advances in Neural Information Processing Sys- tems, vol. 32 (2019). https://doi.org/10.48550/arXiv.1905.02249 . https://proceedings.neurips.cc/paper/2019/hash/1cd138d0499a68f4bb72bee04bbec2d7- Abstract.html

    Pith/arXiv arXiv doi:10.48550/arxiv.1905.02249 2019

  34. [34]

    Xu, L., Ma, S., Zhang, H., Lu, Y.: BAAI General Embeddings (BGE): A General- purpose Text Embedding Model (2023)

    work page 2023

  35. [35]

    https://api.semanticscholar.org/CorpusID: 9540064

    Jarvis, R.A., Patrick, E.A.: Clustering Using a Similarity Measure Based on Shared Near Neighbors (1973). https://api.semanticscholar.org/CorpusID: 9540064

    work page 1973

  36. [36]

    https://arxiv.org/abs/2507.05201

    Sellergren, A., Kazemzadeh, S., Jaroensri, T., Kiraly, A., Traverse, M., Kohlberger, T., Xu, S., Jamil, F., Hughes, C., Lau, C., Chen, J., Mahvar, F., Yatziv, L., Chen, T., Sterling, B., Baby, S.A., Baby, S.M., Lai, J., Schmidgall, S., Yang, L., Chen, K., Bjornsson, P., Reddy, S., Brush, R., Philbrick, K., Asiedu, M., Mezerreg, I., Hu, H., Yang, H., Tiwar...

    Pith/arXiv arXiv 2025

  37. [37]

    Note: The full list has been submitted as the supplementary material

    Douze, M., Guzhva, A., Deng, C., Johnson, J., Szilvasy, G., Mazar´ e, P.-E., Lomeli, M., Hosseini, L., J´ egou, H.: The faiss library (2025) arXiv:2401.08281 [cs.LG] 16 A Appendix : Supplementary Table T able 2: An excerpt of women’s health guidelines and publications used to prepare the questions corpus in this research. Note: The full list has been subm...

    Pith/arXiv arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.