{"id":"5ff14f56-7f04-465a-b068-6d850eb190ee","arxiv_id":"2502.09227","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A research plan proposes pairing neural networks with ILP systems so that AI predictions come with human-readable logical rules, with early tests in weather, law, and biology.","lead":"This paper is a doctoral research statement outlining a plan to combine deep learning with inductive logic programming tools, ILASP and FastLAS, to generate explainable predictions in weather forecasting, legal reasoning, and spermatozoa image classification. It reports only preliminary results and contains no completed method or evaluation.","discovery_kind":"unclear","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No faithfulness test links the learned ILP rules to the neural network's own decision boundary, leaving the central explainability claim untested.","rationale":"The paper is explicitly a research plan: Section 5 describes 'initial findings' and says the image-recognition work 'is still at the beginning stages,' so the reader's UNVERDICTED verdict is appropriate. The concern raised here is not merely that experiments are missing; it is that the proposed architecture has no mechanism that couples the neural network's decisions to the ILP learner. The network is trained on data labels, and the ILP system is also trained on data labels, so the learned rules are a second, independent model of the data. Section 4's phrase 'use them to explain the outcome of the RNN model' asserts a relationship that the pipeline does not establish. This is an internal gap between the stated goal and the described method, not a disagreement with external consensus. The concrete synthetic test proposed above would settle whether the gap can be closed by a simple change, such as training the ILP system on network predictions, or whether the core promise of faithful post-hoc explanations requires a fundamentally different design. Under either outcome, the current manuscript provides insufficient evidence to accept the central claim, so the verdict remains UNVERDICTED and no change to the reader's assessment is needed.","tokens_in":10184,"tokens_out":4072,"duration_ms":40469,"concrete_test":"Run a synthetic experiment with a known ground-truth rule, e.g., event = 1 iff rain > threshold and pressure < threshold, with noisy data. Train the planned neural network and FastLAS on the same data. Then compute (i) agreement between the FastLAS rules and the network's predicted labels on a held-out set, and (ii) the rules learned by a second FastLAS run trained on the network's predicted labels instead of the true labels. If the data-learned rules do not match the network's predictions substantially better than a trivial baseline, or if the network-label rules diverge materially from the data-label rules, the Section 4 post-hoc pipeline is not faithful. Report both agreement metrics explicitly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 commits to 'Post-hoc Explainability': FastLAS will learn rules from meteorological data and then 'use them to explain the outcome of the RNN model.' This is the load-bearing step: for the central claim to hold, the logical rules must be faithful to the neural network, meaning they must track the network's actual decision boundary rather than merely fit the training labels. The paper presents no such check. In Section 5, the only quantitative evidence for FastLAS is that 'it can often reach the same accuracy as the other systems' (SVM, RandomForest, Decision Tree); accuracy on the data does not measure agreement with the neural network. A rule set and a network can both achieve comparable accuracy while disagreeing on many instances, and the rules may describe the data-generating process rather than the network's learned function. The same gap appears in the YOLO/ILASP plan: the 68% YOLO accuracy is reported, but no result shows ILASP rules matching YOLO's classifications. Without a faithfulness metric, such as rule accuracy evaluated on the network's predicted labels or decision-boundary agreement, the explanations could be plausible but unfaithful, which would invalidate the paper's core promise that ILP rules 'substantiate' or 'prove' the network's conclusions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a doctoral-consortium-style research statement that proposes integrating deep learning with Inductive Logic Programming (ILP), specifically the ILASP and FastLAS systems, to provide post-hoc symbolic explanations for neural-network predictions. Three application domains are described: weather forecasting, legal reasoning about Italian law, and automated spermatozoa morphology classification using YOLO networks. The manuscript provides background on ASP, LAS/ILASP, FastLAS, CNNs, RNNs, and YOLO; outlines research goals; and reports preliminary, mostly qualitative status of ongoing work in each domain, including a 68% accuracy figure for a YOLO model and a statement that FastLAS can often reach the same accuracy as SVM, RandomForest, and Decision Tree.","tokens_in":10434,"tokens_out":3450,"duration_ms":33999,"significance":"The paper addresses a relevant open problem in explainable AI: producing symbolic, human-readable justifications for black-box neural predictors. If the proposed framework were realized and validated, it would be a useful contribution to neuro-symbolic XAI. The author correctly identifies a meaningful technical gap — ILP rules potentially explaining the decisions of deep models — and leverages mature tools (ILASP, FastLAS, xASP), which is a sensible starting point. However, as presented, the manuscript is a research agenda rather than a completed study: there are no reproducible experiments, no quantitative evaluation of the explanation quality, and no demonstration that the learned rules actually track the neural networks' decision boundaries. The paper's main strengths are its clear framing of the three domains, its identification of relevant prior systems, and its explicit engagement with the faithfulness problem at a conceptual level. These strengths, however, are not yet backed by the experimental evidence that a journal contribution would require.","major_comments":[{"comment":"The central explainability claim—that ILP rules 'explain the outcome of the RNN model' and 'substantiate the network's conclusions'—is not supported by any faithfulness test. The paper never defines what it means for a learned rule set to be faithful to a neural network, nor does it measure rule-to-network agreement. The only quantitative comparison reported, that 'FastLAS can often reach the same accuracy as the other systems' (SVM, RandomForest, Decision Tree), evaluates the rules on the ground-truth labels, not on the RNN's predicted labels. A rule set can match the network's accuracy while disagreeing with it on most individual instances, so the stated post-hoc explainability promise remains unverified. The paper should include a metric such as rule accuracy evaluated on the network's predicted labels, or decision-boundary agreement, and report the results.","section":"Section 4, Section 5"},{"comment":"The reported results are not reproducible or quantitatively substantiated. For the weather experiments, no dataset size, number of instances, accuracy numbers, standard deviations, or test-set details are given; the description of 10-fold cross-validation with 'four days of training data per fold' is underspecified. For the legal experiments, the claim that 'the evaluation of this model showed that it was successfully able to capture the legal distinctions' is made without any reported accuracy, confusion matrix, or comparison to a baseline. The single concrete number, 'the accuracy of the current model reached 68%' for YOLO, lacks a definition of accuracy (e.g., mAP, classification accuracy, per-class accuracy), test-set size, and variance. Without these details, the paper's central claims cannot be assessed.","section":"Section 5 (weather and legal paragraphs)"},{"comment":"The plan to integrate ILASP with YOLO to 'learn how to identify different morphological characteristics, thereby providing explanations' does not address the grounding problem: YOLO operates on raw pixel features, whereas ILASP learns from symbolic predicates. The paper does not describe what input features or intermediate representations will be provided to ILASP, nor how the learned rules will be related to YOLO's internal feature extraction. This is a load-bearing gap because, without a concrete grounding scheme, the proposed explanation step may describe the data rather than the network's decision process. The manuscript should specify the feature engineering or propose an evaluation protocol that would test the faithfulness of the ILASP rules.","section":"Section 5 (image recognition plan)"}],"minor_comments":[{"comment":"The definition of brave entailment is mis-stated: 'an atom a is bravely entailed by a program P if and only if at least one answer set P contains a' should read 'if and only if at least one answer set of P contains a.' Also, the notation for noisy LAS tasks (T noise vs. T_noise) is inconsistent within the same paragraph.","section":"Section 2.2"},{"comment":"The phrase 'The idea, that was accepted in LPNMR' is awkward; it should be 'The idea was accepted at LPNMR 2024 [18].' There is also a typo 'an view' (should be 'a view') in the organization-of-contribution sentence.","section":"Section 1"},{"comment":"The typo 'applyed' should be 'applied' in the sentence 'where YOLO is applyed for cell and cancer detection.'","section":"Section 3"},{"comment":"Figure 1 and Figure 2 are referenced but not adequately described in the text; Figure 1's 'full explanation of the answer set' is not explained, and Figure 2's example YOLO detection has no associated performance details beyond the later 68% accuracy figure.","section":"Section 5"},{"comment":"Some citations appear tangential (e.g., [10], [17]) and are not clearly integrated into the argument; the distinction between peer-reviewed publications and workshop abstracts or preprints (e.g., [19]) should be clarified when describing 'results achieved.'","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This manuscript reads as a doctoral consortium position paper or extended abstract rather than a completed research article. For such a venue the current content may be acceptable, but for a standard journal the absence of reproducible experiments and a faithfulness evaluation is a blocking issue. The novelty also deserves scrutiny: related work already couples ILP with neural networks (e.g., FFNSL, [9]), so the incremental contribution of the proposed framework needs to be articulated more sharply. If the authors can add a concrete evaluation protocol, report metrics on a defined test set, and provide at least one domain where the learned rules are compared against the network's decisions, the paper could be brought to publishable quality."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a doctoral consortium research statement, not a completed research paper. It outlines a reasonable plan: use ILASP and FastLAS to learn logical rules that explain neural network predictions in weather forecasting, legal reasoning, and spermatozoa image classification. If you read it as a PhD roadmap, it is clear and honest. If you read it as a contribution, there is no new algorithm, no theorem, and no reproducible experiment.\n\nWhat the paper does well is scope the problem and survey the right literature. The author cites prior neuro-symbolic work, including FFNSL and the ILP-for-black-box-preference-learning paper, and does not overstate the current results. The writing is direct, and the legal part points to existing published work. That is deserved credit.\n\nThe soft spot is the one the stress-test flagged, and it is real. The central promise is that ILP rules learned from data can \"substantiate\" or \"explain\" a neural network's conclusions. But the paper never checks whether the rules actually agree with the network's decision boundary. The only comparison for FastLAS is accuracy against SVM, RandomForest, and Decision Tree on the same training data. Two models can have similar accuracy and disagree on many individual instances. The rules might be explaining the data-generating process, not the RNN. Without a faithfulness metric, such as rule accuracy evaluated on the network's predicted labels, the core explainability claim is unverified. For a doctoral statement, this gap is normal at this stage. For a paper it would be load-bearing.\n\nThe other weakness is the absence of numbers. The weather section mentions 10-fold cross-validation but gives no accuracy, no error bars, and no dataset size. The legal evaluation is described qualitatively. The 68% YOLO accuracy has no test-set size or variance. That is consistent with a research statement, but it means there is nothing to referee as a scientific claim.\n\nWho is this for? A supervisor or colleague who wants to see how someone plans to combine ILP and deep learning for XAI, and a reader looking for pointers into the ILP-for-XAI literature. That is useful, but it is not a research contribution. I would not cite it, and I would not send it to a full peer-review track as a paper. If the venue is a doctoral consortium or a workshop, it is fine as is; if it is being considered for a journal or conference, it needs the faithfulness experiment and actual numbers before it deserves referee time.","headline":"A coherent doctoral research plan on neuro-symbolic explainability, but it is a plan, not a paper; the central faithfulness claim is untested.","tokens_in":10939,"tokens_out":2203,"would_cite":false,"duration_ms":22427,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68N17","68T27","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid AI plan uses inductive logic programming to give neural-network predictions human-readable rules that validate them.","keywords":["Inductive Logic Programming","Answer Set Programming","Explainable AI","Deep Learning","ILASP","FastLAS","Neural-Symbolic Integration","Explainability"],"falsifier":"Take the weather or spermatozoa data, train the neural model and the ILP system on the same training folds, then measure disagreement on held-out inputs. If the rules and the network disagree on a substantial share of cases, far beyond the network's own error rate, the claim that the rules substantiate the network's conclusions is falsified; the rules would explain the data, not the model.","tokens_in":9995,"feed_emoji":"🧠","tokens_out":6247,"duration_ms":58018,"temperature":0.7,"pith_summary":"This paper argues for a hybrid architecture in which deep learning does the heavy prediction and inductive logic programming supplies the explanation: after a neural network outputs a decision, ILASP or FastLAS learns answer-set rules from the same data and xASP turns those rules into readable justifications. The goal is to get the accuracy of networks without the black box, in three domains: weather forecasting, Italian criminal-law interpretation, and bull spermatozoa morphology classification. Initial results reported here are treated as evidence that the pipeline is viable: FastLAS often matches the accuracy of SVM, RandomForest, and Decision Tree for rainfall prediction, and a YOLO detector reaches 68% accuracy on spermatozoa. A sympathetic reader would therefore see the contribution as a route toward AI systems whose predictions can be audited by domain experts instead of taken on faith.","feed_headline":"Neural nets get human-readable rules in this hybrid AI plan","feed_subtitle":"Logic programs learned by ILASP and FastLAS explain predictions in weather, law, and image tests.","key_machinery":"The load-bearing object is Learning from Answer Sets (LAS), an ILP setting where a learning task is a tuple of background knowledge, a hypothesis space fixed by mode declarations, and weighted partial interpretations; a solution is a set of rules that covers the examples at minimum penalty. FastLAS solves such tasks with user-defined scoring functions and support for numeric comparisons, while ILASP adds recursive hypotheses through conflict-driven ILP. On the explanation side, eXplainable Answer Set Programming (xASP) converts answer sets of the learned program into directed acyclic graphs that display the dependencies behind a conclusion.","core_discovery":"The central claim is programmatic but concrete: logic rules learned from examples, rather than weights or features inside the network, are the medium through which a hybrid system explains itself. The paper asserts that FastLAS can learn human-readable rules of the form 'past conditions predict future state' for rainfall, reaching accuracy comparable to classical machine-learning baselines; that ILASP can reconstruct the reasoning patterns behind Italian court decisions from encoded statutes and around a hundred precedents; and that a basic YOLO network already reaches 68% accuracy on spermatozoon images. If the program succeeds, every prediction in these domains comes with an answer-set justification produced by a solver rather than by introspection of the network.","pith_inferences":["A testable extension the paper leaves implicit is faithfulness: compare the learned rules against the network's own behavior on the same inputs. If the rules and the network disagree often, the rules explain the dataset, not the model, and the central promise would need to be reworded.","The same hybrid could be run in reverse: any input where the network and the learned rules diverge is a candidate for uncertain or out-of-distribution data, making the rule set a cheap audit layer.","The legal results suggest the pipeline could serve as a consistency checker for case law, not only as an explainer, by automatically flagging contradictory precedents."],"forward_implications":["Weather forecasts from the hybrid system can come with rules like 'given past temperature, humidity, and pressure conditions, rain follows', which meteorologists can read and challenge.","Legal decision support can flag incoherences: ILASP learning from Court of Cassation decisions already surfaced inconsistencies among previous cases.","Spermatozoa morphology classification can be automated, with each abnormal finding accompanied by a logic-based explanation of which morphological criteria triggered the label.","Scoring functions in FastLAS let users bias a learned rule set toward desired properties such as cheapest, safest, or least risky, so explanations can be tuned to the audience.","If FastLAS is made GPU-accelerated, the same explanation pipeline can scale to larger datasets than the current experiments."],"supporting_citations":[{"why":"Foundational definition of inductive logic programming, the paradigm the paper builds on.","marker":"[37]"},{"why":"Introduces FastLAS, the scalable ILP system used for the weather experiments and its scoring functions.","marker":"[30]"},{"why":"Introduces ILASP, the Learning from Answer Sets system used in the legal domain.","marker":"[33]"},{"why":"Supplies xASP, the explanation system that turns answer sets into dependency graphs.","marker":"[2]"},{"why":"Reports the weather-forecasting results with FastLAS that the paper cites as its first concrete evidence.","marker":"[18]"},{"why":"Reports the legal-domain model using ASP and ILASP on Italian statutes and Court of Cassation decisions.","marker":"[12]"},{"why":"Introduces YOLO, the object detector used for spermatozoa classification.","marker":"[40]"},{"why":"Describes conflict-driven ILP, the search strategy ILASP relies on to learn recursive hypotheses.","marker":"[29]"}],"fun_headline_variants":["Logic rules from ILASP explain neural network predictions","Hybrid AI plan uses logic programs for explainability","Deep learning with a logic-based explainer for trust"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the logical rules learned from a dataset truly explain the decisions of the neural network trained on that same dataset, rather than just describing the data in a different way.","fun_headline_variants_meta":{"raw":{"variants":["Logic rules from ILASP explain neural network predictions","Hybrid AI plan uses logic programs for explainability","Deep learning with a logic-based explainer for trust"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"total_tokens":1361,"prompt_tokens":924,"completion_tokens":437,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":388}},"tokens_in":540,"tokens_out":437,"duration_ms":5120,"temperature":1.0,"reasoning_tokens":388,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T22:13:03.525382+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the weather or spermatozoa data, train the neural model and the ILP system on the same training folds, then measure disagreement on held-out inputs. If the rules and the network disagree on a substantial share of cases, far beyond the network's own error rate, the claim that the rules substantiate the network's conclusions is falsified; the rules would explain the data, not the model.","supporting_citations":[{"cited_title":"In: Proc","cited_arxiv_id":null,"evidence_quote":"Introduces FastLAS, the scalable ILP system used for the weather experiments and its scoring functions."},{"cited_title":"The ILASP system for Inductive Learning of Answer Set Programs","cited_arxiv_id":"2005.00904","evidence_quote":"Introduces ILASP, the Learning from Answer Sets system used in the legal domain."},{"cited_title":"Trieu, Tran Cao Son & Marcello Balduccini (2023): Advancements in xASP, an XAI System for Answer Set Programming","cited_arxiv_id":null,"evidence_quote":"Supplies xASP, the explanation system that turns answer sets into dependency graphs."},{"cited_title":"In: Logic Programming and Nonmonotonic Reasoning - 17th International Conference, LPNMR 2024, Dallas, Texas, USA, October 11-14, 2024, Proceedings, Springer","cited_arxiv_id":null,"evidence_quote":"Reports the weather-forecasting results with FastLAS that the paper cites as its first concrete evidence."},{"cited_title":"In: Proceedings of the 39th Italian Conference on Computational Logic, Rome, Italy, June 26-28, 2024 , CEUR Workshop Proceedings 3733, CEUR-WS.org","cited_arxiv_id":null,"evidence_quote":"Reports the legal-domain model using ASP and ILASP on Italian statutes and Court of Cassation decisions."},{"cited_title":"Conflict-driven Inductive Logic Programming","cited_arxiv_id":"2101.00058","evidence_quote":"Describes conflict-driven ILP, the search strategy ILASP relies on to learn recursive hypotheses."}],"review_version":1}