Pith. sign in

REVIEW 3 major objections 6 minor 56 references

An AI certificate can be made into a reproducible statistical claim about performance on a defined application domain.

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 ↗

T0 review

2026-08-04 22:50 UTC pith:KNCB5ZTO

load-bearing objection A transparent audit catalog for EU AI Act compliance; the new value is in the practical audit lessons and the retraining FWER discussion, but the SADD's interpretative flexibility keeps the statistical guarantee conditional. the 3 major comments →

arxiv 2509.08852 v1 pith:KNCB5ZTO submitted 2025-09-08 cs.CY cs.AIcs.LG

Safe and Certifiable AI Systems: Concepts, Challenges, and Lessons Learned

Kajetan Schweighofer , Barbara Brune , Lukas Gruber , Simon Schmid , Alexander Aufreiter , Andreas Gruber , Thomas Doms , Sebastian Eder
show 6 more authors
Florian Mayer Xaver-Paul Stadlbauer Christoph Schwald Werner Zellinger Bernhard Nessler Sepp Hochreiter
This is my paper
classification cs.CY cs.AIcs.LG
keywords functional trustworthinessStochastic Application Domain DefinitionAI certificationEU AI Actminimum performance requirementsstatistical testingdata leakagedistribution shift
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 paper argues that an AI system can be certified as trustworthy only if its performance is checked against a precisely specified population: the application domain, defined as a stochastic sampling recipe, risk-based minimum performance requirements, and a statistical test on independently sampled data. It presents this 'functional trustworthiness' principle as the core of an audit catalog that translates the EU AI Act's obligations into testable criteria for data, model development, robustness, fairness, and monitoring. If the principle is right, a certificate is not a subjective opinion but a reproducible claim that a system meets its stated performance requirements with a controlled error probability. The paper also reports practical failure modes—data leakage, inadequate domain definitions, neglected biases, missing drift controls—that the audit procedure is designed to catch.

Core claim

The central claim is that functional trustworthiness—not general robustness or generic safety—should anchor AI certification. Functional trustworthiness couples three components: a Stochastic Application Domain Definition (SADD) that turns the intended use into a reference distribution p(x) through an explicit sampling strategy; minimum performance requirements derived from risk analysis and desired quality characteristics; and a valid statistical test of those requirements on independent samples drawn according to the SADD. The paper claims this coupling makes performance evidence transparent and reproducible, allows third parties to replicate the evaluation, and gives the certificate's gua

What carries the argument

The Stochastic Application Domain Definition (SADD) carries the argument. It is a three-part recipe: (1) description of the data-generating process and operating context, (2) technical requirements for valid inputs, and (3) a sampling strategy that specifies the reference distribution for evaluation. The SADD matters because without it the reported performance metric has no well-defined population, so the statistical test has nothing to test against and the certificate cannot be reproduced.

Load-bearing premise

The certificate's validity assumes that the written application-domain definition is precise enough that different informed people agree on what counts as a valid test sample; the paper concedes that the SADD is text and cannot be fully formalized.

What would settle it

Have several competent auditors, given the same SADD and the same deployed system, independently draw a test sample according to their reading of the sampling strategy. If their measured performance estimates differ by more than sampling variability would allow, the certificate's guarantees are not reproducible.

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

Share X LinkedIn Reddit HN

If this is right

  • If the framework is correct, a certified system's performance claim is a statistical claim: the system exceeds its minimum performance requirements at a predefined family-wise significance level, not merely on one benchmark.
  • Test data 'representativeness' becomes operational: a test set is representative when it is an independent random sample per the SADD, giving Article 10(3) of the EU AI Act a concrete interpretation.
  • Multiple performance requirements on the same dataset require multiple-testing corrections such as the Bonferroni correction or closed testing procedures; otherwise the certificate's error control is void.
  • A retrained or updated model cannot inherit its predecessor's certificate; sequential procedures such as fixed-sequence testing or the fallback procedure would be needed to control error rates across update cycles.
  • Distribution-shift monitoring can substitute for label-hungry performance monitoring, because the maximum possible performance drop under a domain shift is bounded by the total-variation distance between the original and shifted distributions.

Where Pith is reading between the lines

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

  • The SADD's reliance on the 'average informed user' suggests a testable quality gate: before certifying, an auditor could measure inter-rater agreement among several informed users who independently write sampling strategies from the same SADD; low agreement would expose an ambiguous domain definition.
  • The same statistical core could be extended to post-market surveillance: continuous monitoring with sequential testing would let certificates be maintained across updates while controlling cumulative error, an extension the paper lists as future work.
  • For foundation models and general-purpose AI, which lack a fixed application domain, the SADD approach would have to be inverted—defining a minimal set of operating conditions under which the model can be certified—rather than a single reference population.
  • The adversarial-robustness assessment could be made quantitative in the same minimum-performance-requirement style by setting an explicit maximum perturbation budget and testing whether the model's accuracy under that budget exceeds a threshold, mirroring the framework's treatment of accuracy.

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

3 major / 6 minor

Summary. This white paper presents the TÜV AUSTRIA Trusted AI audit catalog, a conformity-assessment methodology for machine-learning systems aligned with the EU AI Act. The framework rests on three pillars: secure software development, functional requirements, and ethics/data privacy. The core methodological claim is 'functional trustworthiness': an AI system's application domain is defined statistically through a Stochastic Application Domain Definition (SADD), minimum performance requirements (MPRs) are set from risk analysis and legal obligations, and MPRs are verified by statistical tests on independently sampled data. The paper then surveys key certification topics—data leakage, robustness, OOD detection, adversarial examples, distribution shift, uncertainty, bias, fairness, explainability—and post-certification monitoring and retraining, including FWER control via fixed-sequence and fallback procedures. It also reports 'typical mistakes' from the authors' audit practice.

Significance. If the framework holds, this is one of the few concrete operationalizations of AI Act conformity assessment, and it usefully translates legal obligations into testable statistical criteria. The statistical tools used (binomial test, Bonferroni correction, closed testing procedures) are standard and correctly applied in the worked examples. The paper is honest about its limits: Section 3.2 states it is not an accredited scheme, and Section 5.1.4 concedes the SADD cannot be fully formalized. However, the central guarantee is conditional on the SADD being an unambiguous reference distribution, and the paper's own caveat shows this is not achieved. The paper also contains no empirical evidence for its 'lessons learned,' so its value is primarily as a practitioner-oriented methodology description.

major comments (3)
  1. [§5.1.4 (with §4.1 and §5.1.1)] Central claim in §4.1 is statistical guarantees of expected performance, defined with respect to the reference distribution p(x) fixed by the SADD. §5.1.4 concedes the SADD 'is usually provided as a text and cannot be fully formalized' and invokes the 'average informed user' to define the 'true' application domain. Different informed users can operationalize the same text differently, yielding different sampling distributions; §5.1.5 tests then control type I error only for whichever distribution the auditor happens to use. Thus the certificate's guarantee is conditional on a subjective interpretation, not a property of the system in a well-defined domain. The honest admission does not resolve the gap. Please either require a fully operationalized SADD (explicit inclusion/exclusion criteria and a mechanical sampling protocol), or restrict the certificate's claim to the operationalized re
  2. [§5.1.3] The definition of unbiasedness in this central sampling section is incorrect. The text says an estimator is unbiased if it 'reflects the true value across the application domain if it is evaluated on a large enough sample and there are no systematic deviations.' Unbiasedness is a finite-sample property: E[θ̂] = θ for the given sampling design; the 'large enough sample' property is consistency. Please correct the definition and ensure the surrounding discussion of variance and estimation follows from standard definitions.
  3. [§6.2] The fixed-sequence testing section states that 'The sequence terminates as soon as a hypothesis cannot be rejected, as all subsequent hypotheses will also be non-rejectable.' This is a procedural rule of the fixed-sequence scheme, not a statistical property of the hypotheses; in the fallback procedure, later hypotheses can still be tested with their allocated weight. As written, the sentence could mislead readers implementing the procedure. Please rephrase and clarify the distinction between the testing algorithm and the underlying FWER control property.
minor comments (6)
  1. [§5.1.2] 'The corresponding significance level can be set even higher for particularly critical elements' should read 'lower' (i.e., stronger evidence); a higher α makes rejection easier, which is presumably not intended for critical MPRs.
  2. [§5.1.5] The table reports a confidence interval [0.885, 1.000] for n=100, which appears to be a one-sided interval, but the text does not say so. Specify whether intervals are one- or two-sided and state the method (e.g., Clopper-Pearson).
  3. [§5.1.6] The formula (1-α)^n for 'the probability of not falsely rejecting at least one' assumes independent tests. Bonferroni-style FWER control does not require independence. Please state the assumption or label the formula as an illustration.
  4. [§5.3.4] Equation (7) writes the epistemic term as I(p(y,w|x,D)); standard notation is I(y; w | x, D), the mutual information between the target and the parameters. Please fix the notation.
  5. [Abstract / §3.2] The paper uses 'certification' freely, but §3.2 correctly states the scheme is not accredited and the certificate is only a conformity statement. This caveat should also appear in the abstract or introduction to avoid overclaiming.
  6. [§5.3.3] The two-stage shift-detection approach applies two-sample tests to features from a learned encoder. The statistical validity of those p-values depends on treating the encoder as fixed; please state the assumptions or use a data-splitting / conditional-testing framework.

Circularity Check

0 steps flagged

No significant circularity: self-citations are contextual, the statistical testing scheme is a standard hypothesis-test framework, and the acknowledged SADD ambiguity is a validity limitation rather than a circular reduction.

full rationale

The paper is a methodology whitepaper, not a derivation. Its central claim — that functional trustworthiness couples a Stochastic Application Domain Definition (SADD), minimum performance requirements (MPRs), and statistical testing to provide statistical guarantees — is presented as a framework, and the statistical testing in Section 5.1.5 is a textbook binomial test (H0: cacc ≤ 0.9 vs H1: cacc > 0.9) with no fitted parameters renamed as predictions. The SADD is defined in Section 5.1.1 and is explicitly aligned with standard sampling theory: Section 5.1.3 states that 'The application domain corresponds to what is termed population in statistical sampling literature.' The paper acknowledges in Section 5.1.4 that 'the SADD is usually provided as a text and cannot be fully formalized in a mathematical sense' and relies on an 'average informed user' to interpret it; this is a serious validity limitation (the reference distribution is not uniquely pinned down), but it is openly stated and does not make the certificate's claim circular — it makes it conditional on human interpretation. Self-citations appear (Nessler et al. 2023 for the functional trustworthiness principle, Zellinger 2020 Theorem 2.1 for the total-variation bound in Eq. (10), Schmid et al. 2024 for SADD background), but each is either a pointer to a concept explained in-line or a standard mathematical theorem used as independent support. Eq. (10) is the textbook identity that the supremum of an expectation gap over [0,1]-valued functions equals total variation distance; citing Zellinger 2020 for it does not make the monitoring conclusion circular. No fitted input is called a prediction, no uniqueness theorem is imported, and no ansatz is smuggled in via citation. The minor self-citations are not load-bearing for the central claim, so the circularity score is low.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The framework rests on the possibility of defining a stable, interpretable application domain (SADD), setting performance thresholds independently of observed results, and relying on cited statistical results. No free parameters are fitted and no new physical or mathematical entities are introduced.

axioms (4)
  • domain assumption Test data can be sampled independently and at random from the application domain as defined in the SADD.
    The statistical guarantee of MPR testing requires the reference distribution to be fully specified and samples drawn from it; Sections 5.1.3 and 5.1.4 rely on this.
  • domain assumption MPRs and the family-wise significance level are fixed a priori and not adapted to observed model performance.
    The paper states this requirement in 5.1.2, but if violated the statistical test becomes circular and the certificate is meaningless.
  • ad hoc to paper The 'average informed user' interpretation of the SADD yields a well-defined application domain.
    Section 5.1.4 admits the SADD cannot be formalized and relies on a legal concept; this interpretive premise is load-bearing for the whole certification approach.
  • standard math Equation (10), claimed from Zellinger 2020 Theorem 2.1, correctly bounds performance decrease by total variation distance for the considered models and distributions.
    The theorem is cited without proof in Section 6.1, and the paper applies it to justify shift monitoring as a proxy for performance monitoring.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Safe and Certifiable AI Systems: Concepts, Challenges, and Lessons Learned." pith.science (2026). https://pith.science/paper/KNCB5ZTO

@misc{pith2026250908852,
  author       = {Pith},
  title        = {Pith review of: Safe and Certifiable AI Systems: Concepts, Challenges, and Lessons Learned},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KNCB5ZTO}},
  note         = {Machine review of arXiv:2509.08852}
}
Share X LinkedIn Reddit HN
read the original abstract

There is an increasing adoption of artificial intelligence in safety-critical applications, yet practical schemes for certifying that AI systems are safe, lawful and socially acceptable remain scarce. This white paper presents the T\"UV AUSTRIA Trusted AI framework an end-to-end audit catalog and methodology for assessing and certifying machine learning systems. The audit catalog has been in continuous development since 2019 in an ongoing collaboration with scientific partners. Building on three pillars - Secure Software Development, Functional Requirements, and Ethics & Data Privacy - the catalog translates the high-level obligations of the EU AI Act into specific, testable criteria. Its core concept of functional trustworthiness couples a statistically defined application domain with risk-based minimum performance requirements and statistical testing on independently sampled data, providing transparent and reproducible evidence of model quality in real-world settings. We provide an overview of the functional requirements that we assess, which are oriented on the lifecycle of an AI system. In addition, we share some lessons learned from the practical application of the audit catalog, highlighting common pitfalls we encountered, such as data leakage scenarios, inadequate domain definitions, neglect of biases, or a lack of distribution drift controls. We further discuss key aspects of certifying AI systems, such as robustness, algorithmic fairness, or post-certification requirements, outlining both our current conclusions and a roadmap for future research. In general, by aligning technical best practices with emerging European standards, the approach offers regulators, providers, and users a practical roadmap for legally compliant, functionally trustworthy, and certifiable AI systems.

Figures

Figures reproduced from arXiv: 2509.08852 by Alexander Aufreiter, Andreas Gruber, Barbara Brune, Bernhard Nessler, Christoph Schwald, Florian Mayer, Kajetan Schweighofer, Lukas Gruber, Sebastian Eder, Sepp Hochreiter, Simon Schmid, Thomas Doms, Werner Zellinger, Xaver-Paul Stadlbauer.

Figure 1
Figure 1. Figure 1: Deep Learning is a subfield of Machine Learning, which in turn is a subfield of Artificial Intelligence. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Common ML models: Support vector machines, decision trees and neural networks. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FCNNs (left), CNNs (middle) and RNNs (right). The kernel of CNNs (dark blue) is shared and applied to the [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Supervised learning: each input has a target which the model aims to predict (left). Different supervised learning [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Unsupervised learning: Generative methods replicate the data-generating sampling process (left); recoding [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The agent-environment in￾teraction loop in RL. In RL, an agent selects actions based on the current state, receives a re￾ward from the environment, and transitions to a new state, see [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: A foundation model is trained on large-scale data, centralizing general knowledge of a domain. The foundation [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Utilizing a foundation model for specific tasks through fine-tuning, in-context learning or retrieval augmenta [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Multi-modal foundation models can process one or multiple input modalities and output them in one or multiple [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Standardization requests by the European Commission to CEN/CENELEC, also see [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The three pillars of theTÜV AUSTRIA Trusted AI audit catalog: Secure Software Development, Functional Requirements, and Ethics & Data Privacy. In the TÜV AUSTRIA Trusted AI framework, the evaluation of trustworthy AI is based on three pillars, shown in [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Ensuring the functional trustworthiness of an AI system by valid statistical testing of the fulfilment of the minimum performance requirements on independently sampled test data from the application domain. 4.2 The Heart of the Audit Catalog: Functional Requirements The functional requirements are the core of the TÜV AUSTRIA Trusted AI certification and deal with the validation of the data and ML models. … view at source ↗
Figure 13
Figure 13. Figure 13: Lifecycle of AI system development and assurance [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comparison of two sampling strategies with different resulting distributions. The performance estimate can [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Interpreting the Stochastic Application Domain Definition (SADD). (a) A single informed user’s interpreta [PITH_FULL_IMAGE:figures/full_fig_p032_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Decision flowchart for statistical testing of performance metrics: The procedure starts by defining the perfor [PITH_FULL_IMAGE:figures/full_fig_p033_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Multiple testing increases the probability of falsely rejecting the null hypothesis and arriving at wrong con [PITH_FULL_IMAGE:figures/full_fig_p034_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Example of data leakage: Consider a model designed to classify the audio of bird calls. Assume the training [PITH_FULL_IMAGE:figures/full_fig_p035_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Different aspects of robustness. The application domain yields datapoints that are either class 1 (blue) or [PITH_FULL_IMAGE:figures/full_fig_p037_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Hopfield Boosting creates a strong learner by sampling and combining a set of weak learners close to the [PITH_FULL_IMAGE:figures/full_fig_p039_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Instructional illustration for an adversarial example and broader picture of adversarial machine learning. [PITH_FULL_IMAGE:figures/full_fig_p040_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Dimensionality reduction and two-sample tests for distribution shift detection, recreated after [PITH_FULL_IMAGE:figures/full_fig_p041_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Example considering a binary classification dataset, represented by blue and gray dots. Machine learning [PITH_FULL_IMAGE:figures/full_fig_p043_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Technical and societal bias. Technical bias corresponds to a systematic deviation from a true value, exemplified [PITH_FULL_IMAGE:figures/full_fig_p044_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Pareto-front of a machine learning system. Pareto-optimal models are on the pareto-front, pareto sub-optimal [PITH_FULL_IMAGE:figures/full_fig_p046_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Left: performance degradation over time. Right: Retraining after a distribution shift. [PITH_FULL_IMAGE:figures/full_fig_p048_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Illustration of the multiple testing problem. Each statistical test increases the the likelihood of incorrectly [PITH_FULL_IMAGE:figures/full_fig_p050_27.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

56 extracted references · 20 canonical work pages · 4 internal anchors

  1. [7]

    Caron, H

    M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, and A. Joulin. Emerging properties in self-supervised vision transformers. ArXiv, 2104.14294,

    Pith/arXiv arXiv

  2. [8]

    Caton and C

    S. Caton and C. Haas. Fairness in Machine Learning: A Survey. ArXiv, 2010.04053,

    Pith/arXiv arXiv 2010

  3. [12]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Under- standing. ArXiv, 1810.04805,

    Pith/arXiv arXiv

  4. [14]

    Dubey, A

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Yang, A. Fan, A. Goyal, A. Hartshorn, A. Yang, A. Mitra, A. Sravankumar, A. Korenev, A. Hinsvark, A. Rao, A. Zhang, and A. R. et al. The Llama 3 Herd of Models. ArXiv, 2407.21783,

    Pith/arXiv arXiv

  5. [16]

    Accessed: 2025-03-25

    URL https://ec.europa.eu/transparency/documents-register/detail?ref=C(2023)3215&lang= en. Accessed: 2025-03-25. European Commission. Commission publishes Guidelines on AI system definition to facilitate the first AI Act’s rules application, 2024a. URL https://digital-strategy.ec.europa.eu/en/library/commission-publishes-guidelines-ai- system-definition-fa...

    work page 2023

  6. [18]

    Accessed: 2025- 03-25

    URL https://www.europeanlawinstitute.eu/fileadmin/ user_upload/p_eli/Publications/ELI_Response_on_the_definition_of_an_AI_System.pdf. Accessed: 2025- 03-25. S. Farquhar, J. Kossen, L. Kuhn, and Y. Gal. Detecting Hallucinations in Large Language Models Using Semantic Entropy. Nature, 630(8017):625–630,

    work page 2025

  7. [25]

    Kleinberg, S

    J. Kleinberg, S. Mullainathan, and M. Raghavan. Inherent trade-offs in the fair determination of risk scores.ArXiv, 1609.05807,

    Pith/arXiv arXiv

  8. [28]

    J. Lin, R. Men, A. Yang, C. Zhou, M. Ding, Y. Zhang, P. Wang, A. Wang, L. Jiang, X. Jia, J. Zhang, J. Zhang, X. Zou, Z. Li, X. Deng, J. Liu, J. Xue, H. Zhou, J. Ma, j. Yu, Y. Li, W. Lin, J. Zhou, J. Tang, and H. Yang. M6: A chinese multimodal pretrainer. ArXiv, 2103.00823,

    Pith/arXiv arXiv

  9. [29]

    F. T. Liu, K. M. Ting, and Z.-H. Zhou. Isolation forest. In 2008 eighth ieee international conference on data mining, pages 413–422. IEEE,

    work page 2008

  10. [30]

    C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, and D. Ha. The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery. ArXiv, 2408.06292,

    Pith/arXiv arXiv

  11. [31]

    Lundberg and S.-I

    S. Lundberg and S.-I. Lee. A Unified Approach to Interpreting Model Predictions. ArXiv, 1705.07874,

    Pith/arXiv arXiv

  12. [34]

    Muennighoff, Z

    N. Muennighoff, Z. Yang, W. Shi, X. L. Li, L. Fei-Fei, H. Hajishirzi, L. Zettlemoyer, P. Liang, E. Candès, and T. Hashimoto. s1: Simple test-time scaling. ArXiv, 2501.19393,

    Pith/arXiv arXiv

  13. [35]

    Functional trustworthiness of AI systems by statistically valid testing

    B. Nessler, T. Doms, and S. Hochreiter. Functional trustworthiness of AI systems by statistically valid testing. ArXiv, 2310.02727,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [37]

    Parliament and the Council of the European Union

    E. Parliament and the Council of the European Union. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending certain Union legislative acts. https: //eur-lex.europa.eu/eli/reg/2024/1689,

    work page 2024

  15. [38]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 8748–8763. ...

    Pith/arXiv arXiv

  16. [39]

    J. W. Rae, S. Borgeaud, T. Cai, K. Millican, J. Hoffmann, F. Song, J. Aslanides, S. Henderson, R. Ring, S. Young, E. Rutherford, T. Hennigan, J. Menick, A. Cassirer, R. Powell, G. vanDenDriessche, L. A. Hendricks, M. Rauh, P.-S. Huang, a. Glaese, J. Welbl, S. Dathathri, S. Huang, J. Uesato, J. Mellor, I. Higgins, A. Creswell, N. McAleese, A. Wu, E. Elsen,...

    Pith/arXiv arXiv

  17. [40]

    Rajpurkar

    55 Safe and Certifiable AI Systems: Concepts, Challenges, and Lessons Learned P. Rajpurkar. CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning. ArXiv, 1711.5225,

    work page arXiv

  18. [41]

    Ramesh, P

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen. Hierarchical text-conditional image generation with CLIP latents.ArXiv, 2204.06125,

    Pith/arXiv arXiv

  19. [42]

    Why Should I Trust You?

    PMLR. M. T. Ribeiro, S. Singh, and C. Guestrin. "Why Should I Trust You?": Explaining the Predictions of Any Classifier.ArXiv, 1602.04938,

    Pith/arXiv arXiv

  20. [45]

    Y. Sale, V. Bengs, M. Caprio, and E. Hüllermeier. Second-Order Uncertainty Quantification: A Distance-Based Approach. In Forty-first International Conference on Machine Learning (ICML), 2024a. Y. Sale, P. Hofman, T. Löhr, L. Wimmer, T. Nagler, and E. Hüllermeier. Label-wise Aleatoric and Epistemic Uncertainty Quantifi- cation. ArXiv, 2406.02354, 2024b. A....

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [46]

    Shoeybi, M

    M. Shoeybi, M. Patwary, R. Puri, P. LeGresley, J. Casper, and B. Catanzaro. Megatron-LM: training multi-billion parameter language models using model parallelism. ArXiv, 1909.08053,

    Pith/arXiv arXiv 1909

  22. [48]

    Thoppilan, D

    R. Thoppilan, D. deFreitas, J. Hall, N. Shazeer, A. Kulshreshtha, H.-T. Cheng, A. Jin, T. Bos, L. Baker, Y. Du, Y. Li, H. Lee, H. S. Zheng, A. Ghafouri, M. Menegali, Y. Huang, M. Krikun, D. Lepikhin, J. Qin, D. Chen, Y. Xu, Z. Chen, A. Roberts, M. Bosma, V. Zhao, Y. Zhou, C.-C. Chang, I. Krivokon, W. Rusch, M. Pickett, P. Srinivasan, L. Man, K. Meier-Hell...

    Pith/arXiv arXiv

  23. [49]

    Tobin, R

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS),

    work page 2017

  24. [50]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, A. Rodriguez, A. Joulin, E. Grave, and G. Lample. Llama: Open and efficient foundation language models. ArXiv, 2302.13971, 2023a. H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. ...

    Pith/arXiv arXiv

  25. [51]

    X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097–2106,

    work page 2097

  26. [52]

    Y. Wang, Q. Yang, Z. Zeng, L. Ren, L. Liu, B. Peng, H. Cheng, X. He, K. Wang, J. Gao, W. Chen, S. Wang, S. S. Du, and Y. Shen. Reinforcement Learning for Reasoning in Large Language Models with One Training Example. ArXiv, 2504.20571,

    Pith/arXiv arXiv

  27. [54]

    P. M. Winter, S. Eder, J. Weissenböck, C. Schwald, T. Doms, T. Vogt, S. Hochreiter, and B. Nessler. Trusted Artificial Intelligence: Towards Certification of Machine Learning Applications. ArXiv, 2103.16910,

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [55]

    L. Yuan, D. Chen, Y.-L. Chen, N. Codella, X. Dai, J. Gao, H. Hu, X. Huang, B. Li, C. Li, C. Liu, M. Liu, Z. Liu, Y. Lu, Y. Shi, L. Wang, J. Wang, B. Xiao, Z. Xiao, J. Yang, M. Zeng, L. Zhou, and P. Zhang. Florence: A new foundation model for computer vision.ArXiv, 2111.11432,

    Pith/arXiv arXiv

  29. [56]

    Zhang, S

    S. Zhang, S. Roller, N. Goyal, M. Artetxe, M. Chen, S. Chen, C. Dewan, M. Diab, X. Li, X. V. Lin, T. Mihaylov, M. Ott, S. Shleifer, K. Shuster, D. Simig, P. S. Koura, A. Sridhar, T. Wang, and L. Zettlemoyer. OPT: Open pre-trained transformer language models. ArXiv, 2205.01068,

    Pith/arXiv arXiv

  30. [57]

    A. Zou, Z. Wang, N. Carlini, M. Nasr, J. Z. Kolter, and M. Fredrikson. Universal and Transferable Adversarial Attacks on Aligned Language Models. ArXiv, 2307.15043,

    Pith/arXiv arXiv

  31. [1965]

    Weerts, R

    H. Weerts, R. Xenidis, F. Tarissan, H. P. Olsen, and M. Pechenizkiy. Algorithmic Unfairness through the Lens of EU Non- Discrimination Law: Or Why the Law is not a Decision Tree. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’23, page 805–816, New York, NY, USA,

    work page 2023

  32. [1985]

    Conformity assessment – Vocabulary and general principles

    ISO/IEC 17000:2020. Conformity assessment – Vocabulary and general principles. URL https://www.iso.org/standard/ 73029.html. ISO/IEC 25059:2023. Systems and Software Engineering – Systems and software Quality Requirements (SQuaRE) – Quality model for AI systems. URL https://www.iso.org/standard/82570.html. ISO/IEC 27001:2022. Information security, cyberse...

    work page 2020

  33. [1986]

    Soltan, S

    S. Soltan, S. Ananthakrishnan, J. FitzGerald, R. Gupta, W. Hamza, H. Khan, C. Peris, S. Rawls, A. Rosenbaum, A. Rumshisky, C. S. Prakash, M. Sridhar, F. Triefenbach, A. Verma, G. Tur, and P. Natarajan. AlexaTM 20B: Few-shot learning using a large-scale multilingual Seq2Seq model. ArXiv, 2208.01448,

    Pith/arXiv arXiv

  34. [1987]

    M. I. Jordan. Serial order: a parallel distributed processing approach. technical report, june 1985-march

    work page 1985

  35. [1988]

    Mehrabi, F

    N. Mehrabi, F. Morstatter, N. Saxena, K. Lerman, and A. Galstyan. A Survey on Bias and Fairness in Machine Learning. ArXiv, 1908.09635,

    Pith/arXiv arXiv 1908

  36. [1990]

    Esser, S

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, 51 Safe and Certifiable AI Systems: Concepts, Challenges, and Lessons Learned Z. English, K. Lacey, A. Goodwin, Y. Marek, and R. Rombach. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. ArXiv, 2403.03206,

    Pith/arXiv arXiv

  37. [1991]

    Andrychowicz, B

    OpenAI, M. Andrychowicz, B. Baker, M. Chociej, R. Jozefowicz, B. McGrew, J. Pachocki, A. Petron, M. Plappert, G. Powell, A. Ray, J. Schneider, S. Sidor, J. Tobin, P. Welinder, L. Weng, and W. Zaremba. Learning Dexterous In-Hand Manipulation. ArXiv, 1808.00177,

    Pith/arXiv arXiv

  38. [1995]

    Carlini and D

    N. Carlini and D. Wagner. Towards Evaluating the Robustness of Neural Networks. In 2017 IEEE Symposium on Security and Privacy (SP), pages 39–57, 2017a. N. Carlini and D. Wagner. Adversarial Examples Are Not Easily Detected: Bypassing Ten Detection Methods. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security , AISec ’17, page 3...

    work page 2017

  39. [1997]

    Hochreiter and J

    S. Hochreiter and J. Schmidhuber. Nonlinear ICA through low-complexity autoencoders. InProceedings of the 1999 IEEE International Symposium on Circuits and Systems (ISCAS’99), volume 5, pages 53–56. IEEE, 1999a. S. Hochreiter and J. Schmidhuber. Feature extraction through LOCOCODE. Neural Computation, 11(3):679–714, 1999b. J. Hoffmann, S. Borgeaud, A. Men...

    Pith/arXiv arXiv 1999

  40. [2005]

    Matejka and G

    J. Matejka and G. Fitzmaurice. Same stats, different graphs: generating datasets with varied appearance and identical statistics through simulated annealing. In Proceedings of the 2017 CHI conference on human factors in computing systems, pages 1290–1294,

    work page 2017

  41. [2006]

    N. Du, Y. Huang, A. M. Dai, S. Tong, D. Lepikhin, Y. Xu, M. Krikun, Y. Zhou, A. W. Yu, O. Firat, B. Zoph, L. Fedus, M. Bosma, Z. Zhou, T. Wang, Y. E. Wang, K. Webster, M. Pellat, K. Robinson, K. Meier-Hellstern, T. Duke, L. Dixon, K. Zhang, Q. V. Le, Y. Wu, Z. Chen, and C. Cui. GLaM: efficient scaling of language models with mixture-of-experts. ArXiv, 2112.06905,

    Pith/arXiv arXiv

  42. [2007]

    Kotelevskii and M

    N. Kotelevskii and M. Panov. Predictive Uncertainty Quantification via Risk Decompositions for Strictly Proper Scoring Rules. ArXiv, 2402.10727,

    Pith/arXiv arXiv

  43. [2009]

    Y. Sale, P. Hofman, L. Wimmer, E. Hüllermeier, and T. Nagler. Second-Order Uncertainty Quantification: Variance-Based Measures. ArXiv, 2401.00276,

    work page internal anchor Pith review Pith/arXiv arXiv

  44. [2012]

    Denouden, R

    T. Denouden, R. Salay, K. Czarnecki, V. Abdelzad, B. Phan, and S. Vernekar. Improving reconstruction autoencoder out-of- distribution detection with mahalanobis distance. ArXiv, 1812.02765,

    Pith/arXiv arXiv

  45. [2014]

    I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. ArXiv, 1412.6572,

    Pith/arXiv arXiv

  46. [2015]

    K. Lee, H. Lee, K. Lee, and J. Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. ArXiv, 1711.09325,

    Pith/arXiv arXiv

  47. [2016]

    Riccio, F

    P. Riccio, F. Galati, K. Schweighofer, N. Garcia, and N. Oliver. ImageSet2Text: Describing Sets of Images through Text. ArXiv, 2503.19361,

    work page arXiv

  48. [2017]

    Chowdhery, S

    A. Chowdhery, S. Narang, J. Devlin, M. Bosma, G. Mishra, A. Roberts, P. Barham, H. W. Chung, C. Sutton, S. Gehrmann, P. Schuh, K. Shi, S. Tsvyashchenko, J. Maynez, A. Rao, P. Barnes, Y. Tay, N. Shazeer, V. Prabhakaran, E. Reif, N. Du, B. Hutchinson, R. Pope, J. Bradbury, J. Austin, M. Isard, G. Gur-Ari, P. Yin, T. Duke, A. Levskaya, S. Ghemawat, S. Dev, H...

    Pith/arXiv arXiv

  49. [2018]

    A. Auer, P. Podest, D. Klotz, S. Böck, G. Klambauer, and S. Hochreiter. TiRex: zero-shot forecasting across long and short horizons with enhanced in-context learning. ArXiv, 2505.23719,

    arXiv

  50. [2019]

    URL https://www.technologyreview.com/2019/01/ 53 Safe and Certifiable AI Systems: Concepts, Challenges, and Lessons Learned 21/137783/algorithms-criminal-justice-ai/ . A. Kendall and Y. Gal. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? In Advances in Neural Information Processing Systems, volume

    work page 2019

  51. [2020]

    T. B. Brown., B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amo...

    Pith/arXiv arXiv 2005

  52. [2021]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    DeepSeek-AI. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. ArXiv, 2501.12948,

    Pith/arXiv arXiv

  53. [2022]

    Hendrycks and K

    D. Hendrycks and K. Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. ArXiv, 1610.02136,

    Pith/arXiv arXiv

  54. [2023]

    USENIX Association. N. Carlini, M. Jagielski, C. A. Choquette-Choo, D. Paleka, W. Pearce, H. Anderson, A. Terzis, K. Thomas, and F. Tramèr. Poisoning Web-Scale Training Datasets is Practical. In 2024 IEEE Symposium on Security and Privacy (SP),

    work page 2024

  55. [2024]

    Bommasani, D

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill, et al. On the opportunities and risks of foundation models. ArXiv, 2108.07258,

    Pith/arXiv arXiv

  56. [2025]

    Accessed: 2025-03-24

    URL https://digital-strategy.ec.europa.eu/en/library/third-draft- general-purpose-ai-code-practice-published-written-independent-experts . Accessed: 2025-03-24. European Law Institute. Response to the Commission Guidelines on the Application of the Definition of an AI System and the Prohibited AI Practices Established in the AI Act,

    work page 2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.