From Large Language Model Predicates to Logic Tensor Networks: Neurosymbolic Offer Validation in Regulated Procurement

Cedric Haufe; Frieder Stolzenburg

REVIEW 3 major objections 5 minor 19 references

From Large Language Model Predicates to Logic Tensor Networks: Neurosymbolic Offer Validation in Regulated Procurement

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A language model extracts predicate truth values from procurement documents, and a logic tensor network aggregates them with hand-coded fuzzy rules to make offer validity decisions that carry explicit evidence trails.

desk verdict A credible neurosymbolic proof-of-concept for procurement offer validation, but the explanation layer is disconnected from the actual decision logic, so the core XAI claim needs rework. read the letter →

arxiv 2604.05539 v2 pith:GUOZLB2J submitted 2026-04-07 cs.AI

Cedric Haufe , Frieder Stolzenburg This is my paper · ORCID

classification cs.AI

keywords neurosymbolicAIlogictensornetworksexplainableprocurementoffervalidationlargelanguagemodelsfuzzydocumentclassification

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 procurement offer validation in regulated institutions can be made auditable by combining a language model with a logic tensor network. The language model extracts truth values for eight domain-specific predicates from chunks of each document; the logic tensor network aggregates these values with six hand-coded fuzzy rules to decide whether the document is a valid offer. The pipeline also outputs the predicate values, per-rule truth values, and the text passages that drove each decision. Experiments on 200 real offer documents show F1 scores comparable to neural and rule-based baselines, with the authors' claimed advantage being interpretability and modularity rather than raw accuracy. The authors present the work as a proof of concept, noting the small, single-annotator, German-language corpus and hand-coded domain rules.

What carries the argument

The machinery is a two-stage pipeline. Stage one is a predicate layer: a local large language model scores eight semantic predicates (e.g., 'offer title', 'offer number', 'terms of payment') on retrieved text chunks, using one of two self-evaluation schemes (multi-class self-reflection or confidence-based self-consistency). Stage two is an LTN decision layer: the eight predicates are expanded into eleven derived channels (clear vs. vague evidence), passed through learnable gates, and combined by fuzzy logic operators (Gödel, Product, or Łukasiewicz) according to six hand-coded rules that define offer validity. The LTN is trained only to set the gates and decision threshold, not end-to-end, s

What would settle it

Take a held-out batch of procurement documents and have independent procurement experts rate each of the eight predicates; if their ratings disagree with the LLM's predicate scores in cases where those scores change the final decision, the explanation layer is not faithful. A simpler check: delete or add a single rule and show that the decision changes on documents where human experts agree the rule is irrelevant, revealing the rule set is not aligned with true criteria.

Watch

Extended reading notes

Core claim

The central claim is that a decision about whether a procurement document counts as a valid offer can be derived from explicit predicates extracted by a language model and aggregated by a fuzzy rule layer, such that every decision carries a path of evidence: the predicate truth values, the truth values of each applied rule, and the specific text passages that support them. The paper demonstrates this on a real corpus, reporting that the pipeline's classification performance is comparable to a fine-tuned transformer baseline and a plain LLM baseline, while offering explainability those models lack. The authors are careful to state that this is a proof of concept constrained by a small data se

Load-bearing premise

The load-bearing premise is that the eight hand-chosen predicates, the eleven derived channels, and the six fixed rules fully capture what makes an offer valid in this institution, and that the language model's numeric predicate scores faithfully measure those properties — if decisive criteria are missing or the scores are miscalibrated, the explanations are disconnected from the true decision even if the F1 score looks acceptable.

Editorial extensions

If this is right

Offer validation in regulated procurement can be performed with a decision trail that is traceable to individual text passages, which supports legal and institutional review.
The predicate layer is modular: replacing the LLM-based extraction (e.g., with a different estimator or a rule-based extractor) does not require re-designing the decision logic.
Running the full pipeline with locally deployed models keeps sensitive tender documents inside the institution's IT infrastructure, consistent with data-protection requirements.
Rule truth values produced by the LTN can be checked against procurement guidelines, making the system usable for audit or dispute resolution.
If the approach transfers, the same architecture could be re-applied to other regulated document decisions by swapping the predicate set and rules.

Reading between the lines

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

The audit value of the pipeline depends on the completeness of the hand-coded predicate set; a systematic failure-mode analysis of false positives/negatives could reveal missing predicates that human experts would consider decisive.
The observation that a classic LTN baseline outperformed the LLM-based predicate variants in the paper suggests the predicate extraction stage, not the fuzzy logic layer, is the current bottleneck on predictive accuracy — an inference the authors do not make explicitly.
A natural extension is a human-in-the-loop setup where a procurement officer reviews only those documents whose rule truth values fall near the decision threshold, focusing human attention where the model is least certain.
The same two-stage architecture could be tested on other regulated document types (invoices, compliance forms, grant applications) to see whether the claimed interpretability advantage generalizes beyond the single institution and language.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

The paper does a solid job assembling LLM predicate extraction with LTN aggregation and testing it on a real 200-document corpus. The classification results are plausible: their best MCSR+LTN variant lands around 0.87 F1, close to BERT (0.86) and a larger LLM (0.88). That part of the story holds up. But the key advantage they claim — auditable decisions via rule truth values — does not hold as written, because the final decision is not computed from the rules R1–R6. It comes from a gated formula, Obase = PosCore ∧ ¬NegCore, and the rules are only reported afterward. That makes the explanations post-hoc. Under Łukasiewicz implication, R1 (Tc→Obase) is automatically true whenever Tc ≤ Obase, regardless of whether the title actually influenced the decision. So rule truth values can be high even when the decision is driven by unrelated channels. The paper's own claim about "explicit support for XAI" is therefore misleading in its current form.

What is genuinely new here is the application: using LLM-derived predicates and a differentiable fuzzy logic layer for offer validation in regulated procurement. The 11 derived channels are a sensible way to separate clear from vague evidence, and the comparison of MCSR vs. CISC predicate estimation under different fuzzy backends is systematic. The authors are also candid about the small, single-annotator corpus and the domain-specific nature of the rules. I credit them for that.

Beyond the explanation disconnect, there are a few softer problems. Table 1 reports the post-hoc best backend per pipeline on test folds, which inflates the headline scores. There is no significance testing across the 25 folds. No code or data are provided, so I can't independently verify the numbers. And the predicates themselves are not validated against any ground truth other than the final label.

The central empirical comparison is not circular, and the paper is worth engaging with. The fix is clear: either compute the final decision using the actual rules, or explicitly say the rules are illustrative summaries and then evaluate how well they track the real decision. A serious referee could push the authors to make the explanation layer honest.

I would send this to peer review. It's a narrow proof-of-concept, but the problem is real and the gap between claimed and actual explainability is fixable.

Referee Report

3 major / 5 minor

Summary. The paper proposes a neurosymbolic pipeline for validating procurement offers in a regulated German public-university setting. An LLM (Qwen2.5-14B/32B, locally deployed) extracts soft truth values for eight domain-specific predicates, which are expanded into 11 channels, and a Logic Tensor Network (LTN) with hand-coded fuzzy rules aggregates them into a binary IS_VALID_OFFER decision. The authors evaluate on N=200 real documents using repeated stratified 5-fold cross-validation (25 folds), comparing two LLM extraction variants (MCSR and CISC) with a classic LTN, BERT, direct LLM, and deterministic IE+rule baselines. Reported positive-class F1 scores are roughly 0.78–0.90, and the paper claims the main advantage is interpretability, modular predicate extraction, and explicit support for XAI through predicate values, rule truth values, and text passages.

Significance. The application domain is timely and the design choices are pragmatic: real procurement documents, local LLMs to respect GDPR constraints, evidence-chunk retrieval, and a concrete LTN-based decision layer. The paper's strengths include repeated stratified cross-validation, several baselines, and candid acknowledgment of limitations (small N, single annotator, German-only corpus, hand-coded rules). If the empirical and interpretability claims are made sound, the pipeline would be a useful practical contribution to auditable document validation. However, as written, the central XAI claim is not yet established: the rules reported as explanations are not used to compute the final decision, and the headline performance numbers are based on a post-hoc selection of the fuzzy-logic backend on test folds. These issues are load-bearing for the paper's stated key advantage and need to be resolved.

major comments (3)

[§5.1 / Table 1] Table 1 is a 'post-hoc summary of the best F1 results' for each pipeline, meaning the fuzzy-logic backend with the highest mean test-fold F1 was selected per pipeline. This selection is made on the test folds, so the headline numbers (e.g., 0.899 for LTN, 0.874 for MCSR-BestConf) are optimistically biased and do not support a fair comparison across pipelines. The abstract's 'performance comparable to existing models' claim rests on these selected numbers. Since full results are already in Table 2, the authors should either choose the backend on training folds via nested validation, report an average across backends, or explicitly treat backend choice as a tuned hyperparameter and quantify the selection bias. Without this, RQ3 and the overall performance comparison are not conclusive.
[§3.3, Eqs. (2)–(7)] The final decision is computed as Obase(d) = PosCore(d) ∧ ¬NegCore(d), with PosCore formed from gated predicate channels and NegCore from the NOT channels. The rules R1–R6 in Eqs. (2)–(7) are evaluated and 'output together with the predicate values as part of the explanation', but they are not used to compute Obase or the final decision. Thus the reported rule truth values are post-hoc diagnostics rather than the decision logic. For example, under Łukasiewicz implication, R1 (Tc→Obase) is automatically true whenever Tc ≤ Obase, regardless of whether OFFER_TITLE actually contributed to the decision. The learned gates can further reweight channels without changing the rule formulas. This structural disconnect undermines the paper's central claim that decisions 'can be justified by predicate values, rule truth values, and corresponding text passages.' I recommend either incorporating R1–R6
[§2.2 / §2.3] The interpretability claim also depends on the predicates and rules being faithful to the institution's actual validity criteria. The paper provides no predicate-level evaluation against human annotation, no miscalibration analysis for the LLM scores, and no check that the eight predicates plus R1–R6 are complete. Given the small N=200 and single annotator, the LLM's numeric scores could be systematically wrong even if the final F1 looks acceptable. The authors should report per-predicate agreement (e.g., thresholded accuracy or F1 against a held-out human-annotated subset) and a leave-one-predicate-out sensitivity analysis. That would also make the generated explanations substantially more credible.

minor comments (5)

[§3.3] The notation is confusing: p(d) is defined as an 8-dimensional vector, but the implementation uses an 11-dimensional derived vector p̃(d). Table 3 should explicitly state which of the 11 channels come from which of the 8 predicates so that Eqs. (1)–(7) are unambiguous.
[§4.2] The 'IE + deterministic Rules' baseline is described only as 'pattern recognition' followed by a formula. For a fair baseline, the patterns and scoring formula should be specified or cited; otherwise readers cannot assess whether this is a strong deterministic reference.
[§5.5] The error analysis is anecdotal ('False positives predominantly originate from invoice-like documents...'). Quantifying this claim (e.g., counts or proportions of error types) would strengthen the robustness discussion.
[Table 2 caption] Table 2's caption says 'The F1 score of the positive class is given,' but the table also reports precision, recall, and accuracy. The caption should be corrected.
[§6 / abstract] The conclusion states that the pipeline 'enables ... making the decision taken explainable.' Given major comment 2, this is too strong as written; the wording should be softened unless the decision logic is changed.

Circularity Check

1 steps flagged · score 6.0 of 10

The XAI layer is circular: rule truth values are post-hoc functions of the Obase score they are supposed to justify, though the held-out F1 comparison is independent.

self definitional [Section 3.3, Table 4 (Eqs. 1–7), 'Explicit offer rules' and 'Final Decision and Training']
"R1(d) : Tc(d) → Obase(d), (2) ... R4(d) : PosFeature(d) → Obase(d), (5) ... These rule truth values can later be output together with the predicate values as part of the explanation of a decision ... The LTN decision layer outputs a base evaluation Obase(d). ... In our implementation, this base score is the decision on how valid an offer is. IS_VALID_OFFER is decided by applying a threshold to the base score."

Each explanation rule has Obase (or ¬Obase) as its consequent, while Obase is the exact score thresholded to produce IS_VALID_OFFER. Obase is defined separately as PosCore ∧ ¬NegCore from the gated channels; it is not computed by aggregating R1–R6. Thus every reported rule truth value is a function of the decision variable it is supposed to explain: e.g., under Łukasiewicz implication, R1 = min(1, 1−Tc+Obase), which is 1 whenever Tc ≤ Obase. The rules cannot be false independently of the output; they restate it. The key XAI claim—decisions justified by rule truth values—therefore reduces by construction to the output score, even though the held-out classification result is an independent empirical measurement.

full rationale

The classification comparison itself is not circular: the F1 scores are computed on held-out folds, the threshold/gates are tuned per training fold, and the LTN, BERT, and LLM baselines are external. No self-citation chain is load-bearing; the LTN and LLM references are standard external work. The circularity is confined to the explainability layer. The rules in Table 4 are not used to compute Obase; they are evaluated after the fact with Obase in the consequent, so their truth values are determined by the decision score they claim to justify. The paper's own conclusion candidly lists limitations (small single-annotator dataset, hand-coded domain-specific rules, German-only), which bear on external validity but do not create circularity. Because the abstract's key advantage is 'interpretability ... and explicit support for XAI', and that support is a post-hoc restatement of the decision, the partial circularity is central enough to score 6.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim does not introduce new physical entities; the 11 derived input channels are a hand-designed feature expansion, not an independently evidenced ontological commitment. The result depends chiefly on the hand-defined predicate/rule set, the single-annotator labels, the faithfulness of LLM predicate estimates, and the assumption that repeated cross-validation folds provide valid uncertainty estimates.

free parameters (3)

LLM predicate gating weights (α_i, 11 channels) = Not reported; trained per fold with BCE
Section 3.3: gates g_i = σ(α_i) scale each derived channel; these weights are learned on training data and directly determine Obase.
F1-optimized decision threshold per fold = Not reported; chosen on training split per fold
Section 3.3: threshold determined by optimizing positive-class F1 on the training split of each fold.
Fuzzy-logic backend choice per pipeline = Łukasiewicz / Product / Gödel (selected as best test-F1 per pipeline)
Section 5.2 and Table 1: the reported F1 is the best over fuzzy backends, chosen on test folds.

assumptions (6)

domain assumption The 8 hand-defined predicates and 11 derived channels are sufficient to characterize valid offers in this procurement context.
Section 2.2: predicates identified by manual pre-review of original offers and internal procurement guidelines; no completeness validation is provided.
domain assumption Manual labels by the first author (partly with budget-department staff) are treated as ground truth.
Section 2.3: annotation performed by the first author; no inter-annotator agreement or independent legal verification.
domain assumption LLM predicate scores p_k(d) faithfully represent the document properties.
Section 3.2: no independent evaluation of predicate accuracy is reported; only final classification performance is measured.
domain assumption Repeated stratified 5-fold CV folds are treated as independent for means and standard deviations.
Section 4.3: 5 repetitions × 5 folds yields 25 folds sharing documents across repetitions, so fold estimates are not independent.
standard math Fuzzy logic T-norm definitions (Gödel, Product, Łukasiewicz) from Hájek are used as background.
Section 3.3 defines the T-norms and cites Hájek [8]; these are standard background mathematics.
ad hoc to paper The hand-coded rules R1-R6 encode the domain's offer-validity logic.
Section 3.3 and Table 4: the rules are domain-specific and hand-coded for this institution; the authors acknowledge this limitation in Section 6.

how reviews work

0 comments

Cite this review

Pith. "Pith review of From Large Language Model Predicates to Logic Tensor Networks: Neurosymbolic Offer Validation in Regulated Procurement." pith.science (2026). https://pith.science/paper/GUOZLB2J

@misc{pith2026260405539,
  author       = {Pith},
  title        = {Pith review of: From Large Language Model Predicates to Logic Tensor Networks: Neurosymbolic Offer Validation in Regulated Procurement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GUOZLB2J}},
  note         = {Machine review of arXiv:2604.05539}
}

read the original abstract

We present a neurosymbolic approach, i.e. combine symbolic and subsymbolic artificial intelligence, to validating offer documents in regulated public institutions. We employ a language model to extract information and then aggregate it with an LTN (Logic Tensor Network) to make an auditable decision. In regulated public institutions, decisions must be made in a manner that is both factually correct and legally verifiable. Our neurosymbolic approach allows existing domain-specific knowledge to be linked to the semantic text understanding of language models. The decisions resulting from our pipeline can be justified by predicate values, rule truth values, and corresponding text passages. Our experiments on a real corpus show that the proposed pipeline achieves performance comparable to existing models, but its key advantage lies in its interpretability, modular predicate extraction, and explicit support for XAI (Explainable AI).

Figures

Figures reproduced from arXiv: 2604.05539 by the authors.

**Figure 1.** Anonymised example of a valid offer document from the corpus. Each document d was manually assigned a binary label IS_VALID_OFFER(d) ∈ {0, 1} which identifies the extent to which it is valid or invalid. A valid offer is marked with IS_VALID_OFFER = 1 and could be used in the real process. Documents that clearly serve a different purpose (e.g. invoices, delivery notes, order confirmations, general price lists, inter… view at source ↗

**Figure 2.** Overview of the pipeline: Incoming potential offers are segmented, evaluated by an LLM, and then used by an LTN to make the final decision. Since these potential offers usually contain several pages and numerous layout elements (text, tables, diagrams), the LLM does not work directly with the entire raw text. Instead, this raw text is divided into a series of meaningful text excerpts (chunks). For this purpose, it … view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 2 linked inside Pith

[1]

Artificial Intelligence303, 103649 (2022)

Badreddine, S., d’Avila Garcez, A., Serafini, L., Spranger, M.: Logic tensor net- works. Artificial Intelligence303, 103649 (2022). https://doi.org/10.1016/j.artint. 2021.103649

arXiv 2022
[2]

Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., Hui, B., Ji, L., Li, M., Lin, J., Lin, R., Liu, D., Liu, G., Lu, C., Lu, K., Ma, J., Men, R., Ren, X., Ren, X., Tan, C., Tan, S., Tu, J., Wang, P., Wang, S., Wang, W., Wu, S., Xu, B., Xu, J., Yang, Yang, H., Yang, J., Yang, S., Yao, Y., Yu, B., Yuan, H., Yuan, Z.,...

arXiv 2023
[3]

Neural-symbolic learning and reasoning: A survey and interpretation

Besold, T.R., d’Avila Garcez, A., Bader, S., Bowman, H., Domingos, P., Hitzler, P., Kühnberger, K.U., Lamb, L.C., Lima, P.M.V., de Penning, L., Pinkas, G., Poon, H., Zaverucha, G.: Chapter 1. Neural-symbolic learning and reasoning: A survey and interpretation. In: Hitzler, P., Sarker, M.K. (eds.) Neuro-symbolic ar- tificial intelligence: the state of the ...

2022
[4]

In: Cao, T., Das, A., Kumarage, T., Wan, Y., Krishna, S., Mehrabi, N., Dhamala, J., Ra- makrishna, A., Galystan, A., Kumar, A., Gupta, R., Chang, K.W

Bodhwani, U., Ling, Y., Dong, S., Feng, Y., Li, H., Goyal, A.: A calibrated reflection approach for enhancing confidence estimation in LLMs. In: Cao, T., Das, A., Kumarage, T., Wan, Y., Krishna, S., Mehrabi, N., Dhamala, J., Ra- makrishna, A., Galystan, A., Kumar, A., Gupta, R., Chang, K.W. (eds.) Pro- ceedings of the 5th Workshop on Trustworthy NLP (Trus...

2025
[5]

In: Burstein, J., Doran, C., Solorio, T

Devlin,J.,Chang,M.W.,Lee,K.,Toutanova,K.:Bert:Pre-trainingofdeepbidirec- tional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the 14 C. Haufe and F. Stolzenburg Association for Computational Linguistics: Human Language Technologies, Vol- ume1(Lon...

doi:10.18653/v1/n19-1423 2019
[6]

In: Sierra, C

Donadello, I., Serafini, L., d’Avila Garcez, A.: Logic tensor networks for semantic image interpretation. In: Sierra, C. (ed.) International Joint Conferences on Arti- ficial Intelligence (IJCAI 2017). pp. 1596–1602. Curran Associates Inc, Red Hook, NY (2017). https://doi.org/10.24963/ijcai.2017/221

doi:10.24963/ijcai.2017/221 2017
[7]

Artificial Intelligence Review56(11), 12387–12406 (2023)

Garcez, A.d., Lamb, L.C.: Neurosymbolic AI: the 3rd wave. Artificial Intelligence Review56(11), 12387–12406 (2023). https://doi.org/10.1007/s10462-023-10448-w

doi:10.1007/s10462-023-10448-w 2023
[8]

Hájek, P.: Metamathematics of fuzzy logic, Trends in logic, vol. 4. Kluwer Aca- demic, Dordrecht (1998). https://doi.org/10.1007/978-94-011-5300-3

doi:10.1007/978-94-011-5300-3 1998

Show all 19 references

[9]

In: Proceedings of the 14th International Joint Conference on Artificial Intelligence – Volume 2

Kohavi, R.: A study of cross-validation and bootstrap for accuracy estimation and model selection. In: Proceedings of the 14th International Joint Conference on Artificial Intelligence – Volume 2. pp. 1137–1143. IJCAI’95, Morgan Kaufmann Publishers Inc, San Francisco, CA, USA ...

1995 doi
[10]

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., Riedel, S., Kiela, D.: Retrieval-augmented generation for knowledge-intensive nlp tasks. In: H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, H. Lin ...

2020 doi
[11]

Synthesis Lectures on Human Language Technologies, Springer In- ternational Publishing and Imprint Springer, Cham, 1st ed

Lin, J., Nogueira, R., Yates, A.: Pretrained Transformers for Text Ranking: BERT and Beyond. Synthesis Lectures on Human Language Technologies, Springer In- ternational Publishing and Imprint Springer, Cham, 1st ed. 2022 edn. (2022). https://doi.org/10.1007/978-3-031-02181-7

2022 doi
[12]

International Journal of Machine Learn- ing Technology 2:1 (2011) (2011), https://arxiv.org/pdf/2010.16061

Powers, D.M.W.: Evaluation: from precision, recall and F-measure to ROC, in- formedness, markedness and correlation. International Journal of Machine Learn- ing Technology 2:1 (2011) (2011), https://arxiv.org/pdf/2010.16061

2011 arXiv
[13]

Digital Society3(1), 1 (2023)

Richmond, K.M., Muddamsetty, S.M., Gammeltoft-Hansen, T., Olsen, H.P., Moes- lund, T.B.: Explainable AI and law: An evidential survey. Digital Society3(1), 1 (2023). https://doi.org/10.1007/s44206-023-00081-z

2023 doi
[14]

Foundations and Trends® in Information Retrieval3(4), 333–389 (2009)

Robertson, S., Zaragoza, H.: The probabilistic relevance framework: BM25 and beyond. Foundations and Trends® in Information Retrieval3(4), 333–389 (2009). https://doi.org/10.1561/1500000019

2009 doi
[15]

Nature machine intelligence1(5), 206–215 (2019)

Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence1(5), 206–215 (2019). https://doi.org/10.1038/s42256-019-0048-x

2019 doi
[16]

In: Adorni, G., Cagnoni, S., Gori, M., Maratea, M

Serafini, L., d’Avila Garcez, A.S.: Learning and reasoning with logic tensor net- works. In: Adorni, G., Cagnoni, S., Gori, M., Maratea, M. (eds.) AI*IA 2016: ad- vances in artificial intelligence, Lecture Notes in Artificial Intelligence, vol. 10037, pp. 334–348. Springer, Ch...

2016 doi
[17]

Information Processing & Management45(4), 427–437 (2009)

Sokolova, M., Lapalme, G.: A systematic analysis of performance measures for classification tasks. Information Processing & Management45(4), 427–437 (2009). https://doi.org/10.1016/j.ipm.2009.03.002

2009 doi
[18]

In: Che, W., Nabende, J., Shutova, E., Pilehvar, M.T

Taubenfeld, A., Sheffer, T., Ofek, E., Feder, A., Goldstein, A., Gekhman, Z., Yona, G.: Confidence improves self-consistency in LLMs. In: Che, W., Nabende, J., Shutova, E., Pilehvar, M.T. (eds.) Findings of the Association for Computational Neurosymbolic Offer Validation in Re...

2025 doi
[411]

https://doi.org/10.18653/v1/2025.trustnlp-main.26

Association for Computational Linguistics, Stroudsburg, PA, USA (2025). https://doi.org/10.18653/v1/2025.trustnlp-main.26

2025 doi

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Artificial Intelligence303, 103649 (2022)

Badreddine, S., d’Avila Garcez, A., Serafini, L., Spranger, M.: Logic tensor net- works. Artificial Intelligence303, 103649 (2022). https://doi.org/10.1016/j.artint. 2021.103649

arXiv 2022

[2] [2]

Bai, J., Bai, S., Chu, Y., Cui, Z., Dang, K., Deng, X., Fan, Y., Ge, W., Han, Y., Huang, F., Hui, B., Ji, L., Li, M., Lin, J., Lin, R., Liu, D., Liu, G., Lu, C., Lu, K., Ma, J., Men, R., Ren, X., Ren, X., Tan, C., Tan, S., Tu, J., Wang, P., Wang, S., Wang, W., Wu, S., Xu, B., Xu, J., Yang, Yang, H., Yang, J., Yang, S., Yao, Y., Yu, B., Yuan, H., Yuan, Z.,...

arXiv 2023

[3] [3]

Neural-symbolic learning and reasoning: A survey and interpretation

Besold, T.R., d’Avila Garcez, A., Bader, S., Bowman, H., Domingos, P., Hitzler, P., Kühnberger, K.U., Lamb, L.C., Lima, P.M.V., de Penning, L., Pinkas, G., Poon, H., Zaverucha, G.: Chapter 1. Neural-symbolic learning and reasoning: A survey and interpretation. In: Hitzler, P., Sarker, M.K. (eds.) Neuro-symbolic ar- tificial intelligence: the state of the ...

2022

[4] [4]

In: Cao, T., Das, A., Kumarage, T., Wan, Y., Krishna, S., Mehrabi, N., Dhamala, J., Ra- makrishna, A., Galystan, A., Kumar, A., Gupta, R., Chang, K.W

Bodhwani, U., Ling, Y., Dong, S., Feng, Y., Li, H., Goyal, A.: A calibrated reflection approach for enhancing confidence estimation in LLMs. In: Cao, T., Das, A., Kumarage, T., Wan, Y., Krishna, S., Mehrabi, N., Dhamala, J., Ra- makrishna, A., Galystan, A., Kumar, A., Gupta, R., Chang, K.W. (eds.) Pro- ceedings of the 5th Workshop on Trustworthy NLP (Trus...

2025

[5] [5]

In: Burstein, J., Doran, C., Solorio, T

Devlin,J.,Chang,M.W.,Lee,K.,Toutanova,K.:Bert:Pre-trainingofdeepbidirec- tional transformers for language understanding. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the 14 C. Haufe and F. Stolzenburg Association for Computational Linguistics: Human Language Technologies, Vol- ume1(Lon...

doi:10.18653/v1/n19-1423 2019

[6] [6]

In: Sierra, C

Donadello, I., Serafini, L., d’Avila Garcez, A.: Logic tensor networks for semantic image interpretation. In: Sierra, C. (ed.) International Joint Conferences on Arti- ficial Intelligence (IJCAI 2017). pp. 1596–1602. Curran Associates Inc, Red Hook, NY (2017). https://doi.org/10.24963/ijcai.2017/221

doi:10.24963/ijcai.2017/221 2017

[7] [7]

Artificial Intelligence Review56(11), 12387–12406 (2023)

Garcez, A.d., Lamb, L.C.: Neurosymbolic AI: the 3rd wave. Artificial Intelligence Review56(11), 12387–12406 (2023). https://doi.org/10.1007/s10462-023-10448-w

doi:10.1007/s10462-023-10448-w 2023

[8] [8]

Hájek, P.: Metamathematics of fuzzy logic, Trends in logic, vol. 4. Kluwer Aca- demic, Dordrecht (1998). https://doi.org/10.1007/978-94-011-5300-3

doi:10.1007/978-94-011-5300-3 1998