Pith. sign in

REVIEW 3 major objections 4 minor 60 references

When Derived Measurements Mislead: Quantifying and Mitigating LLM Over-Trust with Privileged-Modality Reliability Evidence

T0 review · 3 major / 4 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read LLMs can treat upstream-derived measurements as hard facts; a five-metric chain makes that over-trust measurable and partially fixable.

desk verdict Solid evaluation-framework paper: DFOT plus frozen estimands and matched–shuffled controls are the real product; K2 is a modest, honestly scoped baseline. read the letter →

arxiv 2607.28421 v1 pith:NFEC5BLS submitted 2026-07-30 cs.AI

Zongheng Guo , Tao Chen , Tianli Li , Mingzhe Cui , Yang Jiao , Lei Xie , Yi Pan , Xiao Hu
show 1 more author
Manuela Ferrario
This is my paper
classification cs.AI
keywords derived-featureover-trustlargelanguagemodelsreliabilityevidenceprivilegedinformationPPG-ECGmatched-shuffledinterventionmeasurement-to-LLMpipelines
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Modern AI stacks often hand a large language model a compact number or label produced by another model—heart rhythm from a wearable, a diagnosis score, a summary statistic—and the LLM then reasons as if that value were a direct observation. This paper names that failure derived-feature over-trust (DFOT): the downstream model grants an instance-dependent estimate the status of a fact, or uses it outside its valid scope. Using paired wearable PPG and offline ECG as a concrete case, the authors induce two complementary mistakes—accepting a bad rhythm label, and rejecting a good one under misleading history—then score induction, repair, instance-specific evidence use, and unnecessary verification with five fixed-denominator rates. A privileged ECG-to-PPG reliability signal, never shown as ECG to the LLM, improves repair and specificity on a locked patient-held-out test while only slightly raising needless verification. The point is not that one distillation method solves medicine; it is that DFOT is a distinct, testable interface failure and a shared target for better reliability evidence.

What carries the argument

The DFOT metric chain with frozen denominators, especially ESRM: correct repair under matched reliability evidence minus repair under patient-disjoint shuffled evidence, which tests whether revision depends on the current instance rather than generic caution language.

What would settle it

On the same frozen D1/D2 cases, prompts, and parsers, a reliability generator that beats the privileged baseline on matched CRR and ESRM in both challenges with paired intervals still excluding zero—or a control where matched and shuffled evidence yield the same repair, collapsing ESRM—would confirm or break the framework’s discriminative claim.

Watch

Extended reading notes

Core claim

Derived-feature over-trust is a distinct downstream failure—treating an upstream estimate as a direct fact or using it out of scope—and it can be induced, quantified, and partially mitigated with five linked estimands (COTR, CIR, CRR, ESRM, UHR) plus matched-versus-patient-disjoint-shuffled evidence. In a protocol-locked PPG–ECG study, a privileged-distillation reliability baseline lifts the four repair and specificity endpoints by 1.82–6.69 percentage points with all paired intervals excluding zero, while utility harm rises only 0.67 pp.

Load-bearing premise

That offline ECG agreement is a good enough independent judge of when trusting the PPG-derived rhythm is appropriate for each case; if that reference systematically mislabels appropriateness, the rates no longer measure the claimed failure.

Editorial extensions

If this is right

  • Modular LLM pipelines that ingest derived scores need DFOT-style challenge sets, not only upstream accuracy or end-task accuracy.
  • Matched-versus-shuffled evidence packages become a standard test that revision is instance-specific, not mere caution wording.
  • Reliability generators (calibration, ensembles, selective prediction, multimodal teachers) can be compared on the same frozen COTR/CIR/CRR/ESRM/UHR vector.
  • The same induction-and-repair logic can transfer to other derived measurements whenever an independent appropriateness criterion and a withheld higher-fidelity reference exist.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Any stack that turns raw sensors, vision models, or lab pipelines into tokens for an LLM is a natural DFOT host, not only wearables.
  • Near-ceiling D2 repair with larger ESRM gains suggests future work should optimize specificity and low-false-positive operating points, not only average ranking.
  • Separating measurement-use from clinical-status outputs is a reusable interface pattern whenever reliability and diagnosis collide in one verdict field.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper defines derived-feature over-trust (DFOT): a downstream LLM treating an instance-dependent upstream measurement as a direct fact or using it outside its valid scope. Using PPG-derived rhythm with offline ECG as a withheld reference, it induces two failures (D1: accept conflicted PPG; D2: reject reliable PPG under severe history) and proposes five estimands—COTR, CIR, CRR, ESRM, UHR—with frozen denominators and a matched-versus-patient-disjoint-shuffled evidence control. On a protocol-locked 187-patient MIMIC test, an ECG-to-PPG privileged-distillation baseline (K2) improves four repair/specificity endpoints by 1.82–6.69 pp (paired CIs exclude zero) versus a PPG-only reliability baseline (B2), with a small UHR increase. The framework is presented as generator-agnostic; K2 is an illustrative baseline with residual D1 error and unresolved low-FPR behavior.

Significance. If the result holds, DFOT is a useful, auditable interface failure mode between modular estimators and LLM reasoners, distinct from hallucination and upstream miscalibration. The frozen-denominator metric chain, matched–shuffled specificity test, protocol lock, multi-seed hierarchical bootstrap, cross-LLM/prompt checks, negative controls, claim registry, and public code are genuine methodological strengths and make the task comparable for future mitigation work. Even with modest absolute gains and a single physiological case study, the contribution is primarily a reproducible evaluation target rather than a solved clinical system.

major comments (3)
  1. [DFOT: problem formulation and metric chain; Eq. (2)] Problem formulation / Eq. (2) and Cohort and cases: yi, zi, F1/F2, and H are all built from the same four clear-case ECG–PPG rhythm-agreement relations that also structure the ECG teacher and K2 distillation (Eqs. 3–5; Supp. Table 3). Matched–shuffled ESRM isolates package specificity beyond generic reliability language, and B2 vs K2 is a fair generator contrast, but CRR/ESRM still score compliance with that agreement rule rather than a separately adjudicated notion of when PPG use is appropriate. Please state this coupling explicitly as a property of the instantiation (not only as “ECG is not infallible”), and qualify the 1.82–6.69 pp effects as ECG-agreement interface compliance under controlled prompts unless an independent appropriateness criterion is added.
  2. [Results; Table 1; Conclusion] Results / Table 1: D1 matched CRR remains 44.83% under K2 (55.17% unrepaired), D2 CRR is near ceiling, and locked-test low-FPR contrasts are not confirmed (Supp. Table 7). The paper already notes headroom, but the abstract and conclusion still lead with “improves four repair and specificity endpoints.” Balance the lead claim with residual unrepaired burden and the ranking-vs-stringent-operating-point split so readers do not over-read K2 as strong mitigation.
  3. [DFOT evaluation protocol; Supplementary Evidence interface] Benchmark and reporting contract / Evidence interface: the central observables are parser-mapped VERDICT and MEASUREMENT USE fields. Supp. worked examples and zero parse failures help, but there is little validation that these labels track epistemic reliance (vs. prompt-following or lexical sensitivity to HIGH/LOW). A brief sensitivity analysis (alternate phrasings of the evidence block, or human agreement on a sample of verdicts) would strengthen the claim that CRR/ESRM measure reliance repair rather than format compliance.
minor comments (4)
  1. [Figure 2; Table 1] Figure 2 caption refers to validation-stage panels while Table 1 carries the locked E3 replication; make panel–table correspondence explicit in the main text to avoid mixing stages.
  2. [Throughout] Typographical inconsistencies: “W orked metric example,” “F rozen estimands,” “T ransfer beyond,” and “Knoledge Distillation” should be cleaned.
  3. [Abstract; Results; Table 1] UHR’s exploratory 2 pp upper-CI engineering limit is easy to misread as clinical non-inferiority; keep the claim-registry wording in the abstract/results when UHR is first reported.
  4. [Related Work] Related Work is solid on knowledge conflict and LUPI; a short pointer to measurement-error / errors-in-variables decision theory would better situate DFOT outside LLM-only literature.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation circularity: estimands, freeze denominators, and matched–shuffled contrasts are operationally independent of any single reliability fit; shared ECG oracle is criterion design, not a by-construction reduction.

full rationale

The paper’s load-bearing chain is definitional and experimental, not a first-principles derivation that collapses into its inputs. DFOT, D1/D2, and the five estimands (Eq. 2) are fixed from an offline ECG–PPG reference and parser targets before evidence is scored; F1/F2 and H are frozen from no-evidence baselines; ESRM is a matched-versus-patient-disjoint-shuffled contrast that can and does fail under wrong packages. ECG is withheld from the LLM. K2’s privileged KD (Eqs. 3–5) is an illustrative upstream generator evaluated against B2 and global-shuffled teachers; downstream CRR/ESRM gains are empirical paired differences with residual unrepaired mass (e.g., ~55% of D1 errors), not quantities algebraically fixed by the training loss. Self-citations (SIGMA-PPG backbone, related Guo et al. work) supply architecture/init, not uniqueness theorems or forced ansatzes. Sharing one higher-fidelity oracle for zi training labels and for yi/zi evaluation denominators is standard supervised criterion design and a construct-validity caveat, not self-definitional circularity or a fitted parameter renamed as prediction. No step reduces Eq. X to Eq. Y by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 3 invented entities

The central evaluation rests on standard ML/statistics plus domain choices that ECG can adjudicate PPG-derived rhythm reliance offline, that parser verdicts capture over-trust, and that patient-disjoint shuffled evidence packages are the right counterfactual for instance specificity. Free parameters are ordinary training/interface knobs (KD weights, temperature, thresholds) frozen before test access. Invented entities are conceptual/metric constructs (DFOT and the estimands), not physical objects; they have falsifiable handles via the locked challenges.

free parameters (4)
  • KD loss weights (0.15 window + 0.15 global) and temperature T=2 = 0.15 / 0.15, T=2
    Selected on validation and frozen before locked test; enter K2 objective and thus the illustrative baseline’s evidence quality.
  • Language-facing LOW threshold plow ≥ 0.5 = 0.5
    Defines HIGH/LOW evidence labels shown to the LLM; distinct from student-validation operating thresholds.
  • Exploratory UHR interface limit (2 pp upper CI) = 2 percentage points
    Prespecified engineering tolerance for unnecessary verification, not a clinical non-inferiority margin; gates interpretation of harm.
  • Reliability class construction (four clear PPG/ECG rhythm relations) = both-agree HIGH; hidden-irregularity and false-alarm LOW
    Defines zi HIGH/LOW and low-FPR subsets; other relations excluded from clear-label endpoints by design.
assumptions (5)
  • domain assumption Synchronized Lead-II ECG is a higher-fidelity offline rhythm reference sufficient to label appropriate vs inappropriate reliance on PPG-derived rhythm for clear cases.
    Stated in problem formulation and case construction; ECG never claimed infallible but still defines C, N, zi, and targets yi.
  • ad hoc to paper Parser-mapped VERDICT/MEASUREMENT USE fields (Trustworthy/Suspect; Use-As-Is/Verify-First) are valid observables of downstream reliance failures and repairs.
    Required for all five estimands; protocol lessons document earlier collisions when status and reliability shared one label.
  • ad hoc to paper Patient-disjoint shuffled evidence, holding case, baseline answer, template, arm, and seed fixed, isolates evidence-package specificity beyond generic reliability language.
    Defines ESRM; paper notes label and score may change jointly, so factorial isolation of wording is not claimed.
  • domain assumption D1/D2 controlled challenges characterize inducibility and conditional repair, not population prevalence of DFOT.
    Explicit in evaluation protocol; prevents over-reading COTR/CIR as epidemiology.
  • standard math Standard supervised learning, KL distillation, AUROC/bootstrap inference, and temperature-0 LLM decoding behave as usual under the frozen seeds and vLLM settings.
    Background methodology for teacher/student training and paired hierarchical bootstrap.
invented entities (3)
  • Derived-feature over-trust (DFOT) independent evidence
    purpose: Name the downstream failure of assigning instance-dependent derived measurements the epistemic status of direct facts or using them out of scope.
    Core conceptual contribution; operationalized via D1/D2 rather than left purely verbal.
  • Five-estimand metric chain (COTR, CIR, CRR, ESRM, UHR) independent evidence
    purpose: Separate failure induction, frozen-error repair, matched-package specificity, and unnecessary verification harm with fixed denominators.
    Defined in Eq. 2; ESRM and UHR especially are paper-specific constructs for the LLM interface.
  • K2 privileged PPG reliability student (ECG-logit distillation baseline) independent evidence
    purpose: Illustrative deployment-time PPG-only reliability generator to exercise the framework against B2.
    Not required by the DFOT definition; one LUPI-style instantiation among many possible generators.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Derived Measurements Mislead: Quantifying and Mitigating LLM Over-Trust with Privileged-Modality Reliability Evidence." pith.science (2026). https://pith.science/paper/NFEC5BLS

@misc{pith2026260728421,
  author       = {Pith},
  title        = {Pith review of: When Derived Measurements Mislead: Quantifying and Mitigating LLM Over-Trust with Privileged-Modality Reliability Evidence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NFEC5BLS}},
  note         = {Machine review of arXiv:2607.28421}
}
read the original abstract

Derived measurements increasingly enter large language model (LLM) pipelines as direct facts despite their instance-dependent validity. We define derived-feature over-trust (DFOT) as the failure in which a downstream LLM assigns such a measurement the epistemic status of a direct fact or uses it outside its valid scope. Using physiological sensing as a case study, D1 tests acceptance of a PPG-derived rhythm contradicted by offline ECG, whereas D2 tests rejection of an offline-confirmed reliable PPG rhythm under misleading severe history. ECG supplies training supervision and offline reference construction but is never shown to the LLM. Five estimands quantify this chain: conflict over-trust rate (COTR) and context-induced error rate (CIR) characterize D1/D2; correct repair rate (CRR) measures frozen-error repair; evidence-specific repair margin (ESRM) contrasts matched and patient-disjoint shuffled evidence; and utility harm rate (UHR) measures unnecessary verification among HIGH-reliability cases used without verification at baseline. The framework does not depend on a particular reliability generator. We demonstrate it on 50,000 paired PPG-ECG records using ECG-to-PPG privileged distillation as an illustrative baseline and PPG-only inference. On a protocol-locked 187-patient test, the baseline improves four repair and specificity endpoints by 1.82-6.69 percentage points, with all paired confidence intervals excluding zero; UHR increases by 0.67 percentage points (95% CI: -0.4 to +1.7). DFOT provides a common evaluation target for stronger mitigation methods. The code is available at https://github.com/Zongheng-Guo/When-Derived-Measurements-Mislead.

Figures

Figures reproduced from arXiv: 2607.28421 by the authors.

Figure 1
Figure 1. End-to-end instantiation of the DFOT framework. A proof-of-concept privileged student converts training [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. DFOT results in the PPG–ECG case study. (A) Locked-test upstream reliability transfer. (B) Validation￾stage conditional repair and evidence-package specificity. (C) Robustness across LLM families under P1 and across prompt formulations under Qwen3-8B. (D) Interface verification harm. The independent E3 replication is reported in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Exploratory clinician evaluation of the continuous heart-rate DFOT extension. (A) One clinician, blinded to GPT-5.5 ratings and internal analysis labels, scored 50 reports from 25 paired cases on a prespecified 1–5 scale integrating clinical correctness, safety, and actionability. Thin lines connect reports from the same underlying case, and diamonds indicate group means. Bracketed values show 95% case-bootstrap con… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 7 linked inside Pith

  1. [1]

    Physiological Measurement , volume =

    Allen, John , title =. Physiological Measurement , volume =. 2007 , doi =

    2007

  2. [2]

    and others , title =

    Charlton, Peter H. and others , title =. Physiological Measurement , volume =. 2023 , doi =

    2023

  3. [3]

    and Sanchez, Jose M

    Tison, Geoffrey H. and Sanchez, Jose M. and Ballinger, Brandon and others , title =. JAMA Cardiology , volume =. 2018 , doi =

    2018

  4. [4]

    and Mahaffey, Kenneth W

    Perez, Marco V. and Mahaffey, Kenneth W. and Hedlin, Haley and others , title =. New England Journal of Medicine , volume =. 2019 , doi =

    2019

  5. [5]

    International Conference on Learning Representations , year =

    Pillai, Arvind and Spathis, Dimitris and Kawsar, Fahim and Malekzadeh, Mohammad , title =. International Conference on Learning Representations , year =

  6. [6]

    arXiv preprint arXiv:2601.21031 , year =

    Guo, Zongheng and Chen, Tao and Jiao, Yang and Pan, Yi and Hu, Xiao and Ferrario, Manuela , title =. arXiv preprint arXiv:2601.21031 , year =

    arXiv

  7. [7]

    arXiv preprint arXiv:2606.07365 , year =

    Geenjaar, Eloy and Calhoun, Vince and Daly, Scott and others , title =. arXiv preprint arXiv:2606.07365 , year =

    arXiv

  8. [8]

    arXiv preprint arXiv:2509.06516 , year =

    Guo, Zongheng and Chen, Tao and Ferrario, Manuela , title =. arXiv preprint arXiv:2509.06516 , year =

    arXiv

Show all 60 references
  1. [9]

    Journal of the Franklin Institute , volume =

    Guo, Zongheng and Chen, Huahua and Lin, Lili and Zhou, Wenhui and Yang, Meng and Ying, Na and Guo, Chunsheng , title =. Journal of the Franklin Institute , volume =. 2023 , doi =

    2023

  2. [10]

    Biomedical Signal Processing and Control , volume =

    Chen, Huahua and Zhang, Xiang and Guo, Zongheng and Ying, Na and Yang, Meng and Guo, Chunsheng , title =. Biomedical Signal Processing and Control , volume =. 2024 , doi =

    2024

  3. [11]

    Signal Processing , volume =

    Chen, Tao and Cui, Mingzhe and Guo, Zongheng and Wu, Chenhao and Xie, Lei and Mainardi, Luca , title =. Signal Processing , volume =. 2026 , doi =

    2026

  4. [12]

    International Conference on Learning Representations , year =

    Li, Chenqi and Liu, Yu and Denison, Timothy and Zhu, Tingting , title =. International Conference on Learning Representations , year =

  5. [13]

    Biomedical Engineering Letters , volume =

    Tang, Hui and Ma, Gang and Qiu, Lishen and Zheng, Lesong and Bao, Rui and Liu, Jing and Wang, Lirong , title =. Biomedical Engineering Letters , volume =. 2024 , doi =

    2024

  6. [14]

    arXiv preprint arXiv:2510.02410 , year =

    Langer, Patrick and Kaar, Thomas and Rosenblattl, Max and others , title =. arXiv preprint arXiv:2510.02410 , year =

  7. [15]

    arXiv preprint arXiv:1503.02531 , year =

    Hinton, Geoffrey and Vinyals, Oriol and Dean, Jeff , title =. arXiv preprint arXiv:1503.02531 , year =

    arXiv

  8. [16]

    Neural Networks , volume =

    Vapnik, Vladimir and Vashist, Akshay , title =. Neural Networks , volume =. 2009 , doi =

    2009

  9. [17]

    Unifying Distillation and Privileged Information , booktitle =

    Lopez-Paz, David and Bottou, Leon and Sch. Unifying Distillation and Privileged Information , booktitle =

  10. [18]

    , title =

    Guo, Chuan and Pleiss, Geoff and Sun, Yu and Weinberger, Kilian Q. , title =. Proceedings of the 34th International Conference on Machine Learning , pages =

  11. [19]

    Advances in Neural Information Processing Systems , volume =

    Geifman, Yonatan and El-Yaniv, Ran , title =. Advances in Neural Information Processing Systems , volume =

  12. [20]

    Proceedings of the 36th International Conference on Machine Learning , pages =

    Geifman, Yonatan and El-Yaniv, Ran , title =. Proceedings of the 36th International Conference on Machine Learning , pages =

  13. [21]

    Advances in Neural Information Processing Systems , volume =

    Madras, David and Pitassi, Toniann and Zemel, Richard , title =. Advances in Neural Information Processing Systems , volume =

  14. [22]

    , title =

    Kompa, Benjamin and Snoek, Jasper and Beam, Andrew L. , title =. npj Digital Medicine , volume =. 2021 , doi =

    2021

  15. [23]

    Nature , volume =

    Singhal, Karan and Azizi, Shekoofeh and Tu, Tao and others , title =. Nature , volume =. 2023 , doi =

    2023

  16. [24]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =

    Cocchieri, Alessio and Ragazzi, Luca and Tagliavini, Giuseppe and Moro, Gianluca , title =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages =. 2026 , doi =

    2026

  17. [25]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =

    Longpre, Shayne and Perisetla, Kartik and Chen, Anthony and Ramesh, Nikhil and DuBois, Chris and Singh, Sameer , title =. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =. 2021 , doi =

    2021

  18. [26]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages =

    Chen, Hung-Ting and Zhang, Michael and Choi, Eunsol , title =. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages =

    2022

  19. [27]

    arXiv preprint arXiv:2404.10198 , year =

    Wu, Kevin and Wu, Eric and Zou, James , title =. arXiv preprint arXiv:2404.10198 , year =

    arXiv

  20. [28]

    2025 , note =

    Cattan, Arie and Jacovi, Alon and Ram, Ori and Herzig, Jonathan and Aharoni, Roee and Goldshtein, Sasha and Ofek, Eran and Szpektor, Idan and Caciularu, Avi , title =. 2025 , note =

    2025

  21. [29]

    Larger and More Instructable Language Models Become Less Reliable , journal =

    Zhou, Lexin and Schellaert, Wout and Mart. Larger and More Instructable Language Models Become Less Reliable , journal =. 2024 , doi =

    2024

  22. [30]

    arXiv preprint arXiv:2505.09388 , year =

    Yang, An and Li, Anfeng and Yang, Baosong and others , title =. arXiv preprint arXiv:2505.09388 , year =

    arXiv

  23. [31]

    PLOS ONE , volume =

    Saito, Takaya and Rehmsmeier, Marc , title =. PLOS ONE , volume =. 2015 , doi =

    2015

  24. [32]

    and Hinkley, David V

    Davison, Anthony C. and Hinkley, David V. , title =

  25. [33]

    , title =

    Obuchowski, Nancy A. , title =. Biometrics , volume =

  26. [34]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

    Gupta, Saurabh and Hoffman, Judy and Malik, Jitendra , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

  27. [35]

    and Kibbe, Warren A

    Bent, Brinnae and Goldstein, Benjamin A. and Kibbe, Warren A. and Dunn, Jessilyn P. , title =. npj Digital Medicine , volume =. 2020 , doi =

    2020

  28. [36]

    Mikael and Waggott, Daryl and others , title =

    Shcherbina, Anna and Mattsson, C. Mikael and Waggott, Daryl and others , title =. Journal of Personalized Medicine , volume =. 2017 , doi =

    2017

  29. [37]

    , title =

    Torres-Soto, Jessica and Ashley, Euan A. , title =. npj Digital Medicine , volume =. 2020 , doi =

    2020

  30. [38]

    Heart Rhythm O2 , volume =

    Aschbacher, Kirstin and Yilmaz, Defne and Kerem, Yaniv and others , title =. Heart Rhythm O2 , volume =. 2020 , doi =

    2020

  31. [39]

    JACC: Clinical Electrophysiology , volume =

    Antiperovitch, Pavel and Mortara, David and Barrios, Joshua and Avram, Robert and Yee, Kimberly and Khaless, Armeen Namjou and Cristal, Ashley and Tison, Geoffrey and Olgin, Jeffrey , title =. JACC: Clinical Electrophysiology , volume =. 2024 , doi =

    2024

  32. [40]

    and Do, Duc H

    Pereira, Tania and Tran, Nate and Gadhoumi, Kais and Pelter, Michele M. and Do, Duc H. and Lee, Randall J. and Colorado, Rene and Meisel, Karl and Hu, Xiao , title =. npj Digital Medicine , volume =. 2020 , doi =

    2020

  33. [41]

    Advances in Neural Information Processing Systems , volume =

    Lakshminarayanan, Balaji and Pritzel, Alexander and Blundell, Charles , title =. Advances in Neural Information Processing Systems , volume =

  34. [42]

    and Nowozin, Sebastian and Dillon, Joshua and Lakshminarayanan, Balaji and Snoek, Jasper , title =

    Ovadia, Yaniv and Fertig, Emily and Ren, Jie and Nado, Zachary and Sculley, D. and Nowozin, Sebastian and Dillon, Joshua and Lakshminarayanan, Balaji and Snoek, Jasper , title =. Advances in Neural Information Processing Systems , volume =

  35. [43]

    Advances in Neural Information Processing Systems , volume =

    Jiang, Heinrich and Kim, Been and Guan, Melody and Gupta, Maya , title =. Advances in Neural Information Processing Systems , volume =

  36. [44]

    Proceedings of the 37th International Conference on Machine Learning , series =

    Mozannar, Hussein and Sontag, David , title =. Proceedings of the 37th International Conference on Machine Learning , series =

  37. [45]

    Proceedings of the 36th International Conference on Machine Learning , series =

    Raghu, Maithra and Blumer, Katy and Sayres, Rory and Obermeyer, Ziad and Kleinberg, Bobby and Mullainathan, Sendhil and Kleinberg, Jon , title =. Proceedings of the 36th International Conference on Machine Learning , series =

  38. [46]

    , title =

    Jain, Sarthak and Wallace, Byron C. , title =. Proceedings of NAACL-HLT , pages =. 2019 , doi =

    2019

  39. [47]

    Proceedings of EMNLP-IJCNLP , pages =

    Wiegreffe, Sarah and Pinter, Yuval , title =. Proceedings of EMNLP-IJCNLP , pages =. 2019 , doi =

    2019

  40. [48]

    arXiv preprint arXiv:2207.05221 , year =

    Kadavath, Saurav and Conerly, Tom and Askell, Amanda and others , title =. arXiv preprint arXiv:2207.05221 , year =

    arXiv

  41. [49]

    International Conference on Learning Representations , year =

    Kuhn, Lorenz and Gal, Yarin and Farquhar, Sebastian , title =. International Conference on Learning Representations , year =

  42. [50]

    Teacher Privileged Distillation: How to Deal with Imperfect Teachers? , journal =

    Mart. Teacher Privileged Distillation: How to Deal with Imperfect Teachers? , journal =. 2025 , doi =

    2025

  43. [51]

    Yang, Shuo and Sanghavi, Sujay and Rahmanian, Holakou and Bakus, Jan and Vishwanathan, S. V. N. , title =. Advances in Neural Information Processing Systems , volume =

  44. [52]

    IEEE International Conference on Acoustics, Speech and Signal Processing , year =

    Wei, Riling and Chen, Hanjie and Yao, Kelu and Yang, Chuanguang and Wang, Jun and Li, Chao , title =. IEEE International Conference on Acoustics, Speech and Signal Processing , year =

  45. [53]

    arXiv preprint arXiv:2509.19215 , year =

    Ni, Juntong and Kataria, Saurabh and Tang, Shengpu and Yang, Carl and Hu, Xiao and Jin, Wei , title =. arXiv preprint arXiv:2509.19215 , year =

  46. [54]

    Yuan, Yucheng and Ji, Yuanfeng and Li, Zhongxiao and Li, Ruijiang , journal =

  47. [55]

    ICML 2025 Workshop on Programmatic Representations for Agent Learning , year =

    Learning Game-Playing Agents with Generative Code Optimization , author =. ICML 2025 Workshop on Programmatic Representations for Agent Learning , year =

    2025

  48. [56]

    Understanding the Challenges in Iterative Generative Optimization with

    Nie, Allen and Daull, Xavier and Kuang, Zhiyi and Akkiraju, Abhinav and Chaudhuri, Anish and Piasevoli, Max and Rong, Ryan and Yuan, YuCheng and Choudhary, Prerit and Xiao, Shannon and others , journal =. Understanding the Challenges in Iterative Generative Optimization with

  49. [57]

    Communications Engineering , volume=

    Interactive computer-aided diagnosis on medical image using large language models , author=. Communications Engineering , volume=. 2024 , publisher=

    2024

  50. [58]

    arXiv preprint arXiv:2303.04671 , year=

    Visual chatgpt: Talking, drawing and editing with visual foundation models , author=. arXiv preprint arXiv:2303.04671 , year=

    arXiv

  51. [59]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Audiogpt: Understanding and generating speech, music, sound, and talking head , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  52. [60]

    2024 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI) , pages=

    Physiollm: Supporting personalized health insights with wearables and large language models , author=. 2024 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI) , pages=. 2024 , organization=

    2024

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.