Rule-Based Moral Principles for Explaining Uncertainty in Natural Language Generation

Peter R Lewis; Zahra Atf

REVIEW 4 major objections 4 minor 47 references

Rule-Based Moral Principles for Explaining Uncertainty in Natural Language Generation

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Three moral rules replace numeric confidence in LLM answers

desk verdict A well-written vision paper on virtue-based uncertainty explanations, but the evidence is far too thin—and internally inconsistent—to support the claim that moral rules can replace probabilistic confidence. read the letter →

arxiv 2509.07190 v1 pith:QDC332MF submitted 2025-09-08 cs.CL cs.HC

Zahra Atf , Peter R Lewis This is my paper

classification cs.CLcs.HC

keywords uncertaintyexplanationnaturallanguagegenerationvirtueethicsrule-basedsystemPrologtrustcalibrationexplainableAIfairness

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

This paper argues that a large language model does not need calibrated probability scores to handle uncertainty responsibly: three qualitative levels—low, medium, high—can be mapped to moral rules (responsibility, deference, precaution) that decide what the system does and what it says. The rules are encoded in a small Prolog engine, so the mapping from uncertainty tag to action and plain-language rationale is declarative, inspectable, and model-agnostic. On twenty synthetic clinical and legal prompts the engine produced a valid action for every input, though its automatic tagger matched an oracle label only half the time. The authors treat this as evidence that virtue-aligned symbolic rules can replace numerical confidence signals in high-stakes generation, with a user study left as future work.

What carries the argument

The load-bearing mechanism is a three-rule symbolic controller: a Python tagger turns LLM output into low/medium/high uncertainty, and a SWI-Prolog engine carries facts of the form action(Tag, Action) and rationale(Tag, Rationale) that select an ethical response and its explanation. Because the mapping is a small set of declarative clauses, the entire ethical policy can be inspected and altered without touching the language model. The tagger's surface heuristics—variance across completions, token-level log-probability, and epistemic markers like 'might' and 'possibly'—are the only probabilistic residue.

What would settle it

A user study where people receive either numeric confidence scores or the rule-based rationales for the same high-stakes questions, then rate their trust and make a decision, would settle the central claim: if the rationales do not measurably improve calibration or comprehension over raw scores, the framework's benefit over probabilistic methods disappears. Re-running the twenty-prompt evaluation with a tagger that must match or beat a simple entropy baseline would also test whether the 50 percent accuracy is intrinsic to the heuristic approach.

Watch

Extended reading notes

Core claim

The central claim is that a minimal set of virtue-derived rules—precaution, deference, responsibility—can serve as an effective, transparent substitute for probabilistic uncertainty estimates when an LLM must explain itself. High uncertainty triggers warn-and-refer, medium triggers a partial answer with an authoritative reference, and low triggers a full answer with a disclaimer; each action carries a fixed plain-language rationale. The authors implement this mapping in Prolog and evaluate it on twenty clinical and legal scenarios, reporting deterministic coverage, a 50 percent tagging accuracy, a 0.25 fairness gap, and compact rationales. They conclude that these findings validate the thesi

Load-bearing premise

The benefit over numeric confidence rests on the untested premise that a short canned rationale attached to an uncertainty tag genuinely calibrates user trust and improves interpretability; no human participants were tested.

Editorial extensions

If this is right

If the thesis holds, high-stakes LLM systems can signal uncertainty and act on it without exposing raw probabilities, using rationales that are fixed and auditable.
The rule base gives direct traceability to regulatory requirements such as the EU AI Act and NIH trustworthy-AI principles, since each action maps to a specific clause.
Demographic masking at tagging time can remove measured gender disparities in action distribution, suggesting an input-level fairness lever.
Entropy-aware tagging and controlled-language templates are identified as concrete upgrades that improve accuracy and readability without changing the symbolic layer.
Because the engine is model-agnostic and runs in under ten milliseconds per query, it can be deployed on edge devices as a transparent wrapper around existing LLMs.

Reading between the lines

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

If the 50 percent tagger accuracy is not improved, the ethical layer is only as good as the heuristics feeding it: a wrong tag can trigger the wrong virtue, so the framework's moral guarantee is conditional on upstream accuracy.
The untested trust-calibration assumption could be settled by a direct behavioral comparison in which users receive either rule rationales or numeric confidence scores for the same clinical or legal decision.
The same three-rule structure could extend beyond explanation to govern other uncertainty-aware behaviors, such as when to abstain, escalate, or route to a human, and the virtue mapping could be tuned for cultural context.

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.

Desk Editor's Note

read the letter

Quick take: the paper packages a simple Prolog overlay onto LLM uncertainty tags with a virtue-ethics taxonomy. That combination is new, and the authors do a decent job situating it in the philosophical and technical literature. But the central claim—that virtue-aligned rules can 'replace numerical confidence values'—is not supported by what they ran. The only measured performance is 50% tagging accuracy on 20 synthetic prompts, and that's with self-assigned oracle labels. RQ2, which is the whole trust-calibration benefit, is never tested; they explicitly say a user study is needed. On top of that, there's a direct inconsistency: Section IV.A describes a tagger using variance across k=5 completions, but Section IV.C says the evaluation used k=1 and keyword matching. So the 50% figure may not even apply to the architecture they claim to implement. The '64% with entropy' figure is asserted without experiment, and the fairness masking result seems tautological—removing demographic tokens naturally changes keyword-based tags.

What's good: the paper is clearly written and honestly flags several limitations. The Prolog encoding is transparent and lightweight. The mapping from virtues to rules is a nice intellectual hook, and the regulatory-alignment section is a thoughtful touch even if it's just a table. The ablation on hedging/negation cues is a reasonable first step, though on n=20 it's meaningless statistically.

What's not: the evaluation is too small and too shallow. Coverage = 1.00 is trivial because every tag maps to an action. The explanations are canned templates, so the claimed interpretability benefit is unmeasured. The fairness gap of 0.25 disappears when you mask demographics, but that just tells you the tagger keys on demographic tokens—not that the system is fair. And the two mentions of 11% vs 0% in the masked-fairness numbers is another inconsistency.

Who's this for? A reader working on AI ethics, interpretability, or human-AI trust might find the framing worth a skim, but not as a validated method. It's a design proposal with a pilot. If the authors run a proper user study, compare against a probabilistic baseline, fix the k=1/k=5 mismatch, and report confidence intervals on at least a few dozen prompts, the idea could become a solid short paper. As it stands, it's too shaky to cite as evidence. I'd send it to peer review only if the venue is willing to accept a major-revision path, because the core idea is worth engaging with but the current evidence doesn't justify acceptance.

Referee Report

4 major / 4 minor

Summary. This paper proposes a symbolic layer for LLM uncertainty communication. It defines three qualitative uncertainty tags (low, medium, high), maps each via hand-written Prolog facts to an action and a fixed natural-language rationale, and grounds the mapping in virtue ethics (precaution, deference, responsibility). The framework is evaluated on 20 synthetic clinical/legal prompts; Table III reports coverage 1.00, tagging accuracy 0.50, fairness gap 0.25, readability 39.2, and completeness 0.11. The paper claims these results validate the thesis that virtue-aligned rules can replace numerical confidence values, while also presenting qualitative use cases and fairness counterfactuals.

Significance. If the claimed results held, the paper would offer a lightweight, auditable alternative to probabilistic uncertainty estimates in high-stakes NLG, with potential regulatory relevance. The underlying idea is timely, and the paper is transparent about several limitations, includes a small ablation, and provides enough implementation detail to be understood. However, the current evidence is far from sufficient to establish the central claim; the paper is best read as a design proposal with a preliminary simulation rather than a validated framework.

major comments (4)

[§IV.A vs §IV.C] There is a direct inconsistency between the described architecture and the executed evaluation. §IV.A states that the tagger uses variance across k=5 completions, token-level log-probability, and epistemic markers, while §IV.C states that each prompt was processed with k=1 and a Python tagger using heuristic keyword matching. The reported 0.50 tagging accuracy in Table III therefore does not necessarily test the architecture in Fig. 1. This matters because RQ1 concerns the tagger as designed; the evaluation as reported measures a different, weaker configuration. The authors should either align the execution with the described architecture or clearly report both variants and discuss the discrepancy.
[§IV.D/Table III and §V] The central conclusion in §V—'these findings validate the core thesis that virtue-aligned rules can replace numerical confidence values'—is not supported by Table III. Tagging accuracy is 0.50 on n=20 against self-assigned oracle labels; thus, in half of the cases, the triggered moral rule and its rationale are based on the wrong uncertainty level. No comparison is made to probabilistic baselines or to verbalized-confidence methods (e.g., Lin et al. [46]) that the framework is positioned against. The entropy-based prediction in §V ('would raise accuracy to 64%') is explicitly hypothetical and is not a measured result; it cannot count as evidence. At most, the paper demonstrates feasibility of the Prolog layer, not superiority over probabilistic alternatives.
[§V (RQ2)] RQ2 asks whether pairing each rule with a plain-language explanation improves user trust and calibration, but the paper contains no user study or any human-participant data. Section V itself concedes that 'a user study will be needed to verify whether this brevity suffices for informed consent.' All claims about trust calibration and improved interpretability are therefore unsupported. The paper should either present human-subject evidence or clearly label RQ2 as an untested hypothesis and remove it from the list of validated contributions.
[§IV.D/Table III and Listing 2] Coverage 1.00 is a tautology of the implementation: Listing 2 defines action/2 and rationale/2 for exactly the three tags, so every tag necessarily yields an action. It does not measure robustness to noisy completions or 'silent failure' as claimed in §V. Similarly, the 'rationales' are fixed strings determined solely by the tag; they are not generated by moral reasoning in any operational sense. The paper should reframe the evaluation around what is actually measured—tagger accuracy, action-distribution differences, and text statistics—rather than presenting deterministic rule coverage as evidence for the framework's ethical reasoning capacity.

minor comments (4)

[§IV.D] The metric 'Flesch–Kincaid readability score' is ambiguous: 39.2 is a Flesch Reading Ease score, not a grade level. Please specify the scale and, if possible, report both.
[Table VI] The fairness analysis relies on n=20 and confounded prompts (e.g., 'female, 70, chest pain'). The claim that demographic masking 'eliminates' the gender gap should be stated with appropriate caution, since the sample is tiny and the masking effect is likely driven by a small number of instances.
[§IV.A] The end-to-end latency claim ('under 10 milliseconds on a standard laptop') is presented without a measurement protocol or standard deviation. Please provide details of how this was measured and how many runs were averaged.
[References] There are duplicated references (e.g., [8] and [16] are the same Gal & Ghahramani paper) and several arXiv/technical-report entries without complete venue information. A thorough reference cleanup is needed.

Circularity Check

3 steps flagged · score 6.0 of 10

Coverage 1.00 and the fairness-masking result are by-construction consequences of the hand-coded rule base and lexical tagger; the central thesis is validated only via these definitional metrics, yielding partial circularity.

self definitional [Section IV.D (Evaluation Metrics, Coverage), Section IV.C, Listing 2, Table III]
"Coverage— The proportion of prompts for which the reasoning engine successfully produced a valid action label. ... Coverage 1.00"

The tagger in Section IV.C assigns one of low/medium/high by keyword matching; Listing 2 defines action(Tag, Action) and rationale(Tag, Rationale) for exactly those three tags. Therefore every prompt necessarily yields a valid action label, so Coverage=1.00 is a logical consequence of the finite tag domain and exhaustive facts, not an empirical finding. Presenting this as 'flawless coverage' and 'insulating the system against silent failure' converts a definitional property of the hand-written rule base into a measured benefit.
self definitional [Section V (Discussion), with Listings 1-2 and Section III.B]
"Taken together, these findings validate the core thesis that virtue-aligned rules can replace numerical confidence values ... a user study will be needed to verify whether this brevity suffices for informed consent."

The 'findings' used to validate the core thesis include the by-construction Coverage=1.00 and hand-written rationales from Listing 1/2; the delivered explanation is the exact string paired with the tag by the authors. The other numbers (0.50 tagging accuracy, 0.25 fairness gap, 39.2 readability) do not show that moral rules outperform probability scores, and the paper concedes that the trust-calibration premise is untested. The validation statement therefore reduces to the fact that the rule base was defined to cover its own tag space.

1 more flagged steps

self definitional [Section IV.G (Impact of Demographic Masking), Table VII; Section IV.C]
"we generated a masked version of the prompts by replacing all demographic labels (e.g., male, 70) with a generic placeholder (unknown). The tagging pipeline was then re-applied to this masked dataset. As shown in Table VII, the fairness gap (Δ) between male and female groups in the action distribution dropped to zero after masking."

The tagger in Section IV.C assigns tags using heuristic keyword matching on surface text. Masking replaces the demographic tokens with 'unknown', so the tagger cannot use them; the gap dropping to zero is a mechanical consequence of removing the exact features the lexical tagger depends on, not a discovery about the moral layer. The fairness result is thus an artifact of the input transformation by construction.

full rationale

The main circularity is in the evaluation: Coverage=1.00 is guaranteed by the finite tag set and exhaustive action/2 and rationale/2 facts, and the delivered 'moral rationale' is the hand-written string paired with the tag in Listings 1-2, so the 'validation' of the core thesis leans on definitional properties. The fairness-masking result is likewise a consequence of removing the very keywords the lexical tagger uses. Two self-citations ([19], [45]) appear, but they are contextual/motivational and do not carry the derivation; the meta-analysis in [45] is an external published result, so they do not independently raise the score. The paper also contains missing support: RQ2 is untested ('a user study will be needed'), the 64% accuracy and readability-improvement figures in Section V are hypothetical, and the k=1/k=5 architecture inconsistency is unresolved. These are support concerns in addition to the by-construction results; overall, one or more predictions reduce by construction, but the system's tagger accuracy and fairness measurements do contain independent empirical content, so the circularity is partial rather than total.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-chosen mappings, author-assigned labels, and an untested psychological assumption about trust. No empirical constants are derived; the 'virtue-aligned controller' and 'characterful interlocutor' are rhetorical labels for the same lookup table, not new entities.

free parameters (4)

Uncertainty tag thresholds and keyword lists = unspecified
The tagger uses 'surface-level cues' (Section IV-A) with no published cutoffs; the set of epistemic markers (e.g., 'might,' 'possibly') is hand-chosen, and k=5, temperature 0.7 are fixed constants.
Tag to action mapping = high->warn_and_refer, medium->partial_answer_with_reference, low->full_answer_with_disclaimer
Stipulated in Listing 1/2 and Section III-A; no derivation from moral theory beyond assertion.
Oracle uncertainty labels = 3 low, 4 medium, 3 high per domain
Assigned by the authors to 20 synthetic prompts with no inter-annotator agreement or external standard.
Entropy-based accuracy improvement = 64% (claimed)
Section V states a 10-token sliding-window entropy measure 'would raise accuracy to 64%' with no experiment or data.

assumptions (4)

domain assumption Three-level uncertainty taxonomy (low/medium/high) is sufficient to guide ethically appropriate response.
Section I and III define the taxonomy; no evidence that three levels capture the uncertainty relevant to safety.
domain assumption Virtue ethics (Hagendorff's six virtues) is the correct moral framework.
Section III-C cites Hagendorff and Aristotle; the choice among moral theories is assumed, not argued.
ad hoc to paper Author-assigned oracle labels are ground truth.
Section IV-B uses oracle labels for accuracy and fairness; these are the authors' own judgments on synthetic prompts.
ad hoc to paper Plain-language rationales improve trust and interpretability.
Asserted throughout and in RQ2, but no user study is conducted; Section V admits a user study is needed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Rule-Based Moral Principles for Explaining Uncertainty in Natural Language Generation." pith.science (2026). https://pith.science/paper/QDC332MF

@misc{pith2026250907190,
  author       = {Pith},
  title        = {Pith review of: Rule-Based Moral Principles for Explaining Uncertainty in Natural Language Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QDC332MF}},
  note         = {Machine review of arXiv:2509.07190}
}

read the original abstract

Large language models (LLMs) are increasingly used in high-stakes settings, where explaining uncertainty is both technical and ethical. Probabilistic methods are often opaque and misaligned with expectations of transparency. We propose a framework based on rule-based moral principles for handling uncertainty in LLM-generated text. Using insights from moral psychology and virtue ethics, we define rules such as precaution, deference, and responsibility to guide responses under epistemic or aleatoric uncertainty. These rules are encoded in a lightweight Prolog engine, where uncertainty levels (low, medium, high) trigger aligned system actions with plain-language rationales. Scenario-based simulations benchmark rule coverage, fairness, and trust calibration. Use cases in clinical and legal domains illustrate how moral reasoning can improve trust and interpretability. Our approach offers a transparent, lightweight alternative to probabilistic models for socially responsible natural language generation.

Figures

Figures reproduced from arXiv: 2509.07190 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 39 canonical work pages

[46]

Teaching models to express their uncertainty in words,

S. Lin, J. Hilton, and O. Evans, “Teaching models to express their uncertainty in words,”arXiv preprint arXiv:2205.14334, 2022

arXiv 2022
[1]

Assessing LLMs in malicious code deobfuscation of real-world malware campaigns,

C. Patsakis, F. Casino, and N. Lykousas, “Assessing LLMs in malicious code deobfuscation of real-world malware campaigns,”Expert Syst. Appl., vol. 256, 2024

work page 2024
[2]

Large language models versus classical machine learning: Performance in COVID-19 mortality prediction using high-dimensional tabular data,

M. Ghaffarzadeh-Esfahani, A. Salahi-Niri, and H. Toreyhi, “Large language models versus classical machine learning: Performance in COVID-19 mortality prediction using high-dimensional tabular data,” arXiv preprint arXiv:2409.02136, 2024

arXiv 2024
[3]

Large language models as tax attorneys: A case study in legal capabilities emergence,

J. J. Nay, D. Karamardian, S. B. Lawsky, W. Tao, M. Bhat, R. Jain, A. T. Lee, J. H. Choi, and J. Kasai, “Large language models as tax attorneys: A case study in legal capabilities emergence,”Philos. Trans. R. Soc. Lond. A, vol. 382, no. 2270, Apr. 2024

work page 2024
[4]

LegalMind system and the LLM-based legal judgment query system,

A. S. Athul, A. Saxena, J. Mahajan, L. Panikulangara, S. Kulkarni, and S. Bang, “LegalMind system and the LLM-based legal judgment query system,” inProc. Int. Conf. Trends Quantum Comput. Emerg. Bus. Technol. (TQCEBT), 2024

work page 2024
[5]

Can ChatGPT predict Chinese equity premiums?,

F. Ma, Z. Lyu, and H. Li, “Can ChatGPT predict Chinese equity premiums?,”Finance Res. Lett., vol. 65, 2024

work page 2024
[7]

Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,

Y . Gal and Z. Ghahramani, “Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,” inProc. 33rd Int. Conf. Mach. Learn., 2016, pp. 1050–1059

work page 2016
[8]

ChatGPT: A canary in the coal mine or a parrot in the echo chamber? Detecting fraud with LLM: The case of FTX,

G. Gregory and V . Liuzzi, “ChatGPT: A canary in the coal mine or a parrot in the echo chamber? Detecting fraud with LLM: The case of FTX,”Finance Res. Lett., vol. 70, 2024

work page 2024

Show all 47 references

[9]

Language models are few-shot learners,

T. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language models are few-shot learners,”arXiv preprint arXiv:2005.14165, 2020

2005 arXiv
[10]

Anchoring revisited: Robust evidence from large-scale meta-analysis,

X. Zhang and E. H. Brown, “Anchoring revisited: Robust evidence from large-scale meta-analysis,”J. Exp. Psychol. Gen., vol. 152, no. 4, pp. 845–861, 2023

2023
[11]

Resource-rational analysis: Understanding human cognition as the optimal use of limited computational resources,

F. Lieder and T. L. Griffiths, “Resource-rational analysis: Understanding human cognition as the optimal use of limited computational resources,” Behav. Brain Sci., vol. 43, p. e1, 2020

2020
[12]

Natural frequencies illuminate base-rate neglect and facilitate Bayesian reasoning,

R. Hertwig and M. Meder, “Natural frequencies illuminate base-rate neglect and facilitate Bayesian reasoning,”Perspect. Psychol. Sci., vol. 16, no. 1, pp. 3–22, 2021

2021
[13]

Human behaviour in the context of low-probability, high- impact events,

E. Sundh, “Human behaviour in the context of low-probability, high- impact events,”Humanit. Soc. Sci. Commun., vol. 11, no. 1, pp. 1–14, 2024

2024
[14]

Bad at probability? That might be a blessing,

J. Thomson, “Bad at probability? That might be a blessing,”Big Think, Everyday Philosophy column, Feb. 2024

2024
[15]

Simple and scalable predictive uncertainty estimation using deep ensembles,

B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,” inAdv. Neural Inf. Process. Syst., vol. 30, pp. 6402–6413, 2017

2017
[16]

Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,

Y . Gal and Z. Ghahramani, “Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,” inProc. 33rd Int. Conf. Mach. Learn. (ICML), vol. 48, pp. 1050–1059, 2016

2016
[17]

Why should I trust you?: Explaining the predictions of any classifier,

M. T. Ribeiro, S. Singh, and C. Guestrin, “Why should I trust you?: Explaining the predictions of any classifier,” inProc. 22nd ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. (KDD), pp. 1135–1144, 2016

2016
[18]

Man is to computer programmer as woman is to homemaker? Debiasing word embeddings,

T. Bolukbasi, K.-W. Chang, J. Y . Zou, V . Saligrama, and A. T. Kalai, “Man is to computer programmer as woman is to homemaker? Debiasing word embeddings,” inAdv. Neural Inf. Process. Syst., pp. 4349–4357, 2016

2016
[19]

The challenge of uncertainty quantification of large language models in medicine,

Z. Atf, S. A. A. Safavi-Naini, P. R. Lewis, A. Mahjoubfar, N. Naderi, T. R. Savage, and A. Soroush, “The challenge of uncertainty quantification of large language models in medicine,”arXiv preprint arXiv:2504.05278, 2025

2025 arXiv
[20]

Large language model uncertainty proxies: discrimination and calibration for medical diagnosis and treatment,

T. Savage, J. Wang, R. Gallo, A. Boukil, V . Patel, S. A. A. Safavi-Naini, A. Soroush, and J. H. Chen, “Large language model uncertainty proxies: discrimination and calibration for medical diagnosis and treatment,”J. Am. Med. Inform. Assoc., pp. 1–11, 2024

2024
[21]

Uncertainty-aware explainable recommendation with large language models,

C. Lin, J. Hu, H. Guo, B. Kong, S. Hu, X. Wu, and X. Wang, “Uncertainty-aware explainable recommendation with large language models,”arXiv preprint arXiv:2402.03366, 2024

2024 arXiv
[22]

A novel integration strategy for uncertain knowledge in group decision-making with artificial opinions: A DSFIT-SOA-DEMATEL approach,

L. Sheng, Z. Gu, and F. Chang, “A novel integration strategy for uncertain knowledge in group decision-making with artificial opinions: A DSFIT-SOA-DEMATEL approach,” *Expert Syst. Appl.*, vol. 243, pp. 122886, 2024

2024
[23]

Selective prediction-set models with coverage rate guarantees,

J. Feng, A. Sondhi, J. Perry, and N. Simon, “Selective prediction-set models with coverage rate guarantees,” *Biometrics*, vol. 79, no. 2, pp. 811–825, Dec. 2021

2021
[24]

On the foundations of noise-free selective classification,

R. El-Yaniv and Y . Wiener, “On the foundations of noise-free selective classification,” *J. Mach. Learn. Res.*, vol. 11, pp. 1605–1641, 2010

2010
[25]

Holistic evaluation of language models,

P. Liang et al., “Holistic evaluation of language models,” arXiv preprint arXiv:2211.09110, 2023

2023 arXiv
[26]

What uncertainties do we need in Bayesian deep learning for computer vision?,

A. Kendall and Y . Gal, “What uncertainties do we need in Bayesian deep learning for computer vision?,” arXiv preprint arXiv:1703.04977, 2017

2017 arXiv
[27]

Dlugatch, A

R. Dlugatch, A. Georgieva, and A. Kerasidou, “AI-driven decision sup- port systems and epistemic reliance: A qualitative study on obstetricians’ and midwives’ perspectives on integrating AI-driven CTG into clinical decision making,” BMC Med. Ethics, vol. 25, no. 1, pp. 1–11, 2024

2024
[28]

Mechanisms of cancer metastasis,

M. Castaneda, P. den Hollander, N. A. Kuburich, J. M. Rosen, and S. A. Mani, “Mechanisms of cancer metastasis,” Semin. Cancer Biol., vol. 87, pp. 17–31, 2022

2022
[29]

An exploratory survey about using ChatGPT in education, healthcare, and research,

M. Hosseini, C. A. Gao, D. M. Liebovitz, A. M. Carvalho, F. S. Ahmad, Y . Luo, N. MacDonald, K. L. Holmes, and A. Kho, “An exploratory survey about using ChatGPT in education, healthcare, and research,” *PLoS ONE*, vol. 18, no. 10, pp. 1–15, 2023

2023
[30]

Selectively answering ambiguous questions,

J. R. Cole, M. J. Q. Zhang, D. Gillick, J. M. Eisenschlos, B. Dhingra, and J. Eisenstein, “Selectively answering ambiguous questions,” inProc. 2023 Conf. Empir. Methods Nat. Lang. Process. (EMNLP), Singapore, pp. 530–543, 2023

2023
[31]

Selective question answering under domain shift,

A. Kamath, R. Jia, and P. Liang, “Selective question answering under domain shift,” inProc. 58th Annu. Meet. Assoc. Comput. Linguist. (ACL), Online, pp. 5684–5696, 2020

2020
[32]

Calibration of Pre-trained Transformers,

S. Desai and G. Durrett, “Calibration of Pre-trained Transformers,”Proc. Conf. Empir. Methods Nat. Lang. Process. (EMNLP), pp. 295–302, 2020

2020
[33]

COPU: Con- formal Prediction for Uncertainty Quantification in Natural Language Generation,

S. Wang, Y . Jiang, Y . Tang, L. Cheng, and H. Chen, “COPU: Con- formal Prediction for Uncertainty Quantification in Natural Language Generation,”arXiv preprint arXiv:2502.12601, 2025

2025 arXiv
[34]

Logic-LM: Empow- ering large language models with symbolic solvers for faithful logical reasoning,

L. Pan, A. Albalak, X. Wang, and W. Y . Wang, “Logic-LM: Empow- ering large language models with symbolic solvers for faithful logical reasoning,” inFindings Assoc. Comput. Linguist.: EMNLP, Singapore, pp. 3806–3824, 2023

2023
[35]

Deep sym- bolic regression for recurrent sequences,

S. d’Ascoli, P.-A. Kamienny, G. Lample, and F. Charton, “Deep sym- bolic regression for recurrent sequences,” inProc. 39th Int. Conf. Mach. Learn. (ICML), 2022

2022
[36]

A comprehensive survey on neuro-symbolic artificial intelligence,

B. P. Bhuyanet al., “A comprehensive survey on neuro-symbolic artificial intelligence,”J. Artif. Intell. Res., 2024

2024
[37]

Integrating artificial intelligence with the Internet of Things: A review of recent advances,

Z. Luet al., “Integrating artificial intelligence with the Internet of Things: A review of recent advances,”IEEE Internet Things J., 2024

2024
[38]

Historical perspectives on the development of neuro-symbolic reasoning,

B. P. Bhuyanet al., “Historical perspectives on the development of neuro-symbolic reasoning,” inHandbook of Neuro-Symbolic AI, Springer, 2025

2025
[39]

Ghari, M.: A formalization of the Protagoras court paradox in a temporal logic of epistemic and normative reasons.Artificial Intelligence and Law 32, 325–367 (2024)

2024
[40]

McKeon, M.W.: In defense of a normative concept of argument.Argu- mentation38, 247–264 (2024)

2024
[41]

Axiomatizing Conditional Normative Reasoning

Parent, X.: Report on “Axiomatizing Conditional Normative Reasoning”. K¨unstliche Intelligenz38, 107–111 (2024)

2024
[42]

A Virtue-Based Framework to Support Putting AI Ethics into Practice,

T. Hagendorff, “A Virtue-Based Framework to Support Putting AI Ethics into Practice,”Philos. Technol., vol. 35, no. 55, 2022

2022
[43]

Kosourikhina, V ., Handley, S.J.: Conflict detection with invalid infer- ences: All heuristics, no logic.Memory & Cognition(2025)

2025
[44]

Cai, M., Jian, X., Wang, Y ., Li, H.: Concept representation and trust relationship modeling in fuzzy social networks.International Journal of Fuzzy Systems25, 2250–2265 (2023)

2023
[45]

Atf, Z., Lewis, P.R.: Is trust correlated with explainability in AI? A meta-analysis.IEEE Transactions on Technology and Society(2025)

2025
[47]

Greene, Moral tribes: Emotion, reason, and the gap between us and the.m New York, NY , USA: Penguin, 2013

J. Greene, Moral tribes: Emotion, reason, and the gap between us and the.m New York, NY , USA: Penguin, 2013

2013
[48]

The emotional dog and its rational tail: A social intuitionist approach to moral judgment,

J. Haidt, “The emotional dog and its rational tail: A social intuitionist approach to moral judgment,”Psychological Review, vol. 108, no. 4, pp. 814–834, 2001

2001

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [46]

Teaching models to express their uncertainty in words,

S. Lin, J. Hilton, and O. Evans, “Teaching models to express their uncertainty in words,”arXiv preprint arXiv:2205.14334, 2022

arXiv 2022

[2] [1]

Assessing LLMs in malicious code deobfuscation of real-world malware campaigns,

C. Patsakis, F. Casino, and N. Lykousas, “Assessing LLMs in malicious code deobfuscation of real-world malware campaigns,”Expert Syst. Appl., vol. 256, 2024

work page 2024

[3] [2]

Large language models versus classical machine learning: Performance in COVID-19 mortality prediction using high-dimensional tabular data,

M. Ghaffarzadeh-Esfahani, A. Salahi-Niri, and H. Toreyhi, “Large language models versus classical machine learning: Performance in COVID-19 mortality prediction using high-dimensional tabular data,” arXiv preprint arXiv:2409.02136, 2024

arXiv 2024

[4] [3]

Large language models as tax attorneys: A case study in legal capabilities emergence,

J. J. Nay, D. Karamardian, S. B. Lawsky, W. Tao, M. Bhat, R. Jain, A. T. Lee, J. H. Choi, and J. Kasai, “Large language models as tax attorneys: A case study in legal capabilities emergence,”Philos. Trans. R. Soc. Lond. A, vol. 382, no. 2270, Apr. 2024

work page 2024

[5] [4]

LegalMind system and the LLM-based legal judgment query system,

A. S. Athul, A. Saxena, J. Mahajan, L. Panikulangara, S. Kulkarni, and S. Bang, “LegalMind system and the LLM-based legal judgment query system,” inProc. Int. Conf. Trends Quantum Comput. Emerg. Bus. Technol. (TQCEBT), 2024

work page 2024

[6] [5]

Can ChatGPT predict Chinese equity premiums?,

F. Ma, Z. Lyu, and H. Li, “Can ChatGPT predict Chinese equity premiums?,”Finance Res. Lett., vol. 65, 2024

work page 2024

[7] [7]

Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,

Y . Gal and Z. Ghahramani, “Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,” inProc. 33rd Int. Conf. Mach. Learn., 2016, pp. 1050–1059

work page 2016

[8] [8]

ChatGPT: A canary in the coal mine or a parrot in the echo chamber? Detecting fraud with LLM: The case of FTX,

G. Gregory and V . Liuzzi, “ChatGPT: A canary in the coal mine or a parrot in the echo chamber? Detecting fraud with LLM: The case of FTX,”Finance Res. Lett., vol. 70, 2024

work page 2024