{"id":"e01432dd-79d0-4d79-9c0c-686485e1095d","arxiv_id":"2607.15776","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"NeurOWL explains plausible subsumptions in incomplete OWL ontologies by finding missing bridge axioms with embedding ranking and LLM validation.","lead":"NeurOWL combines a logical reasoner, ontology embeddings, and an LLM to decide whether an unprovable class–subclass statement about an incomplete OWL ontology is likely true, and to propose missing axioms that would explain it. Reported F1 scores reach 0.97 on three biomedical/food ontologies, but no code or data is released and the recovery theorem assumes the neural validator it is supposed to prove.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's recovery guarantee is not derived: the proof repeatedly assumes the true bridge concept is top-ranked and LLM-accepted, premises that the stated theorem does not contain.","rationale":"The reader's weakest_assumption identifies exactly the right weak point. The proof of Theorem 1 uses 'can identify' as a step, but identification in NeurOWL is the conjunction of (i) candidate generation, (ii) top-k selection by s̄(C), and (iii) LLM YES/NO verification. The theorem's hypotheses control only (i), so the claimed recovery guarantee is not derived for the implemented system. This is an internal proof gap, not a disagreement with external consensus. I also note that the manuscript says code will be released upon acceptance and that baselines are threshold-tuned on a validation set sampled from the test set, so the empirical claims are plausible but not independently checkable. That supports maintaining a conditional verdict rather than accepting or rejecting outright: the empirical pipeline could still be useful, but the strong formal claim needs either a corrected theorem with explicit neural-success assumptions or evidence that top-k and LLM failures do not occur on the evaluated datasets.","tokens_in":21076,"tokens_out":10352,"duration_ms":120635,"concrete_test":"Replay the proof of Theorem 1 with the stated premises, treating each 'NeurOWL can identify X' occurrence as a proof obligation; check whether the premises imply that X is in the top-k and is accepted by the LLM. If not, the theorem needs an explicit extra assumption. As an empirical probe, instrument NeurOWL to log, for each positive test sample, whether the ground-truth bridge concept is in top-k and LLM-accepted. If any positive sample has a ground-truth bridge that is outside top-k or LLM-rejected, the recovery guarantee cannot be derived from the stated hypotheses alone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central formal claim is Theorem 1: for any subset-and-entailment-minimal solution S to an incomplete normalized EL ontology, iterative applications of NeurOWL recover every axiom in S. The proof in the supplementary material (section 'Poof of Theorem 1') contains an unstated oracle assumption. In the actual system, Stage 3a (Figure 1 / 'Main Structure') ranks all candidates by the embedding score s̄(C) and keeps only top-k; only those are sent to the LLM for verification. The theorem's hypotheses say only that the Stage 3 candidate set contains every concept of the form A, ∃r.A, or A⊓B. They say nothing about rank or LLM acceptance. Yet the proof's base case asserts 'NeurOWL can directly identify the missing axioms in S by identifying A1 in Stage 2a, Stage 2b, or Stage 3a,' and the induction step says 'we can assume NeurOWL identify ∃r.B2 as an intermediate concept in Stage 3a.' Each of these is an additional premise: C can be in the candidate set yet ranked outside top-k (so never queried) or rejected by the LLM despite being semantically valid. In either case the output explanation is not returned, the next iteration is not seeded, and 'every axiom in S' is not recovered. Thus Theorem 1 as stated is not a theorem about the implemented NeurOWL; it describes an idealized variant with a perfect retriever and a perfect LLM. Since the abstract and introduction rely on this theorem to justify the recovery claim, this is the load-bearing weak point.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NeurOWL, a neuro-symbolic framework for reasoning over incomplete OWL ontologies. Given an incomplete ontology O' and a candidate subsumption A⊑B that is not entailed, NeurOWL decides whether the subsumption is semantically plausible and, if so, returns an explanation E consisting of at most two missing axioms such that O' ∪ E ⊢ A⊑B. The pipeline combines a DL reasoner (Stage 1), logical bridging through parents/children (Stage 2), embedding-based candidate retrieval with LLM validation (Stage 3a), and a direct LLM plausibility check (Stage 3b). The authors report experiments on FoodOn, SNOMED CT, and GO under random and hard negative sampling, with F1 scores up to 0.970, and also prove a theorem (Theorem 1) claiming that iterative applications of NeurOWL can recover every axiom in a subset-and-entailment-minimal solution for normalized EL ontologies.","tokens_in":21449,"tokens_out":4350,"duration_ms":48588,"significance":"If the claims are correct, NeurOWL addresses a genuinely useful generalization of TBox abduction: it does not require a predefined hypothesis space and also handles the case where the input subsumption is false. The empirical methodology is reasonable, including ablations, a complex setting with existential restrictions, and case studies. The main theoretical contribution, however, is not established for the implemented system as described, and the empirical support would benefit from statistical rigor. The paper's strengths are its clear system design, the use of real-world ontologies, and the inclusion of explanations in the evaluation metrics (X-F1, X-F1*). The central recovery guarantee is the load-bearing element, and it currently rests on unstated oracle assumptions.","major_comments":[{"comment":"Theorem 1 is not a theorem about the implemented NeurOWL. The theorem's hypotheses only require the Stage 3 candidate set to contain certain concepts; they say nothing about the embedding ranker placing the correct bridge within top-k, nor about the LLM accepting it. Yet the proof repeatedly asserts 'NeurOWL can directly identify the missing axioms' and 'we can assume NeurOWL identify ∃r.B2 ... in Stage 3a'. These are extra oracle assumptions. In the actual Stage 3a, candidates are ranked by s̄(C) and only top-k are sent to the LLM; a true bridge can be ranked outside top-k or rejected by the LLM. The theorem therefore describes an idealized system with a perfect retriever and perfect LLM, not the one evaluated. This directly undermines the abstract's and introduction's claim that iterative applications of NeurOWL 'recover any subset-and-entailment-minimal solutions'. Please either state","section":"Theorem 1 / Supplementary 'Poof of Theorem 1'"},{"comment":"All reported F1, X-F1, and X-F1* values are point estimates without error bars, multiple seeds, or significance tests. Several comparisons are close (e.g., FoodOnA random: NeurOWL OnT 0.960 vs OnT baseline 0.910; SnomedA random: 0.968 vs 0.943). The claim of 'strong and robust performance across different domains' is not fully supported without variance information or at least a repeated-seed analysis. This is a load-bearing issue for the empirical contribution, as the differences could be within noise, especially on the smaller or harder datasets.","section":"Table 1 / Experimental Setting"},{"comment":"There is a notable tension between the retrieval results and the reported F1 scores for Snomed∃. In Table 2, H@1 for Stage 3a is 0.001 (OnT) and H@100 is 0.577; with the default top-k=10, the correct existential bridge is essentially never retrieved. Yet Table 1 reports F1=0.924 for NeurOWL OnT on Snomed∃ random. This implies that the F1 score is dominated by Stage 3b's direct LLM check, not by the bridging mechanism that supposedly provides explanations. The X-F1* of 0.193 confirms that explanations are rarely correct. The paper should explicitly discuss this: the 'success' on F1 does not validate the explanation-generation component, and the reader should not infer that NeurOWL is solving the task as stated for complex bridges.","section":"Table 2 and Table 1 on Snomed∃"},{"comment":"The construction of positive samples for the standard setting may introduce a bias toward easy cases. Positive samples are randomly sampled from all (A,B) such that O⊨A⊑B and O'⊭A⊑B. With only 5% pruning, many such entailments in O are lost due to a single missing axiom, and the bridging concept is often a direct neighbor. The hard-negative setting is more challenging, but the random-negative setting seems to overrepresent simple one-hop cases. Please report the distribution of positive samples by the number of required missing axioms and by the distance between A and B in the taxonomy, so readers can assess the difficulty of the benchmark.","section":"Dataset Construction / Positive samples"}],"minor_comments":[{"comment":"The proof section is titled 'Poof of Theorem 1'; should be 'Proof'. Also, 'obatined' typo in the same section, and 'Snd∃' in Table 2 is inconsistent with 'Snomed∃'.","section":"Supplementary, Proof header"},{"comment":"The labels 'INOR (OnT)' and 'INOR (SBERT)' appear to be typos; presumably should be 'NeurOWL'. This makes the figure confusing.","section":"Figure 2"},{"comment":"The definition of direct child has a typo: 'A is a direct children' should be 'A is a direct child'. Also, the direct-child condition is defined relative to O (the original ontology), while the pruning is on O'; this should be clarified because directness in O' may differ.","section":"Evaluation, footnote 2"},{"comment":"The related work is thorough, but the discussion of Zhao et al. (2026) is very brief; since that work is described as close in using LLMs for plausibility with human verification, a few sentences on the difference would help contextualize NeurOWL.","section":"Related Work"},{"comment":"The scoring function is defined as the average of s(A⊑C) and s(C⊑B), but it is not clarified whether s is the same as the subsumption score used for classification in the baselines. This should be stated explicitly for reproducibility.","section":"System description, Stage 3a"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the disconnect between Theorem 1 and the implemented system. This is fixable: the authors can either state the theorem as an idealization with perfect retrieval and LLM, or restrict the claim to a conditional guarantee, or prove that the top-k and LLM steps are irrelevant under the theorem's minimality conditions (which seems unlikely in general). The empirical work is potentially publishable, but the lack of any variance or significance testing weakens the 'robustness' claim. I would not reject the paper, but I would require a substantive revision that addresses the theorem's scope and the interpretation of the Snomed∃ results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth knowing: NeurOWL gives ontology engineers a practical way to get plausible missing axioms and explanations for incomplete ontologies, and it defines a task (verification plus abduction without pre-selected candidates) that is broader than prior TBox abduction. The paper deserves careful reading. The empirical pipeline is solid and the ablation study is well done. The main formal claim, Theorem 1, is not derived for the actual system.\n\nWhat's new: the task formulation is a genuine generalization of earlier abduction work. Most prior TBox abduction assumes a fixed signature or that the target subsumption is already known to be true. NeurOWL instead asks whether a non-entailed subsumption is plausible and, if so, returns a short explanation of missing axioms. The pipeline itself—reasoner plus embedding-based candidate ranking plus LLM validation—combines existing parts in a sensible way, and the authors are careful to separate retrieval from verification in their analysis. The experiments on FoodOn, GO, and SNOMED CT are extensive, with both atomic and existential-restriction bridge concepts, and the reported F1 scores are strong. The per-stage breakdown and retrieval numbers are useful.\n\nSoft spots, in proportion. The biggest is Theorem 1. The proof repeatedly says things like 'NeurOWL can directly identify the missing axioms ... by identifying A1 in Stage 2a, 2b, or 3a.' But in the actual system, the candidate must survive top-k embedding ranking and LLM validation. The theorem's hypotheses only ensure the relevant concepts are in the candidate set; they say nothing about rank or LLM acceptance. So the guarantee is about an idealized version with perfect retrieval and a perfect LLM, not the implemented NeurOWL. That is a real gap. The abstract and introduction lean on this theorem to claim that iterative applications recover missing axioms, so this is a load-bearing issue. It should be fixable—state the theorem conditionally on retrieval/validation correctness, or verify those properties empirically—but it needs to be fixed.\n\nTwo smaller issues. There is no code or data released, and results are point estimates without error bars; that makes the empirical claims hard to check. And the baselines' thresholds are selected on a validation set sampled from the test set. That is leakage, though it biases against NeurOWL, so it weakens the baselines' numbers rather than NeurOWL's.\n\nWho should read it: anyone working on neuro-symbolic reasoning, LLM-based ontology completion, or abduction in description logics. The empirical contribution and the task framing are worth having, even if the formal claim is overstated.\n\nMy recommendation: send it to peer review. It deserves serious refereeing, but the authors should be asked to fix the theorem, release artifacts, and report variability before the paper is accepted.","headline":"NeurOWL is a genuinely useful pipeline for missing-axiom suggestion and subsumption plausibility, but its recovery guarantee (Theorem 1) is not actually proven for the implemented system.","tokens_in":21944,"tokens_out":4664,"would_cite":true,"duration_ms":47130,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"NeurOWL claims a candidate subsumption can be verified against an incomplete ontology and explained by the missing axioms that would entail it, without preselecting hypotheses.","keywords":["ontology reasoning","subsumption","abduction","neuro-symbolic","LLM","ontology embeddings","EL ontology","explanation"],"falsifier":"Feed NeurOWL a normalized EL ontology with a known minimal solution S, remove one axiom, and query the resulting non-entailed subsumption. The theorem predicts iterative NeurOWL returns exactly S; if repeated runs on such constructed cases return a different or incomplete set while the correct bridge is in the candidate space, the recovery claim is false. A sharper test: pick a case where the ground-truth bridge is in the top-k but the LLM rejects it while accepting a plausible wrong bridge; systematic occurrence of this pattern would falsify the explanation-accuracy claims even with perfect r","tokens_in":20905,"feed_emoji":"🧩","tokens_out":4659,"duration_ms":46063,"temperature":0.7,"pith_summary":"NeurOWL sets out to solve a combined problem: when an ontology does not entail a candidate statement \"A is a subclass of B,\" decide whether the statement is nevertheless plausible, and if it is, return the missing axioms that would make it provable. The framework replaces the usual requirement of a predefined hypothesis space with a search for intermediate \"bridge\" concepts that connect A to B, using ontology embeddings and an LLM to rank and verify candidates. The paper argues that this neuro-symbolic design can both judge plausibility and produce logically sound explanations, and that iterating the procedure can recover any minimal set of missing axioms for normalized EL ontologies. If correct, it gives ontology maintainers a practical way to ask \"should this subsumption hold?\" and receive an actionable, explainable answer.","feed_headline":"One pipeline finds the missing axioms that prove a subsumption","feed_subtitle":"NeurOWL judges whether an is-a statement is plausible and returns the axioms needed to entail it.","key_machinery":"The load-bearing object is the \"bridge concept\" — an intermediate concept C that can connect the left-hand side A to the right-hand side B of a candidate subsumption A⊑B. NeurOWL's pipeline is built around finding such C in three complementary ways: logical bridging over direct children of B and direct parents of A; embedding-based bidirectional retrieval that scores candidate C by the average of s(A⊑C) and s(C⊑B); and a direct LLM check when no bridge exists. The embedding model provides a ranked candidate set, the LLM performs semantic verification, and a reasoner supplies justifications for the non-missing part, so the final explanation is a logically sound combination of new and existing","core_discovery":"On the paper's own terms, the central discovery is that a non-entailed subsumption query can be handled end-to-end by decomposing it into a search for bridge concepts: first look at logical children of B and parents of A (Stages 2a/2b), then use embedding scores to retrieve top-k intermediate concepts C such that A⊑C⊑B (Stage 3a), and if no bridge is found, treat the subsumption itself as a candidate missing axiom and validate it directly with an LLM (Stage 3b). The system returns True with an explanation E iff the LLM validates a bridge, where E is the new axioms along with a justification for the entailed part. Complementing the design, Theorem 1 states that for normalized EL ontologies, i","pith_inferences":["If LLM verification is as reliable as these results suggest, the pipeline could be pointed the other way: given an ontology with known entailments, identify where the LLM disagrees with the reasoner to surface candidate missing axioms for expert review.","The bridge-concept decomposition is not specific to EL; the same retrieve-then-verify skeleton could be adapted to richer description logics, though the recovery guarantee would need a new proof.","The method's dependence on top-k retrieval means explanation quality is capped by retrieval recall; in the complex ∃r.B setting, retrieval H@100 is only 0.577 even with fine-tuning, so the bottleneck has already moved from reasoning to embedding quality.","A natural testable extension is to treat the LLM's YES/NO margin (e.g., via logprobs) as a confidence score and let users trade off precision and recall per domain."],"forward_implications":["Users can query arbitrary subsumptions against an incomplete ontology and receive both a YES/NO plausibility judgment and, on YES, a set of missing axioms that entail the query.","The approach removes the need for a predefined hypothesis space of candidate missing axioms, which prior TBox abduction methods required.","Iterative application lets the system recover a minimal set of missing axioms, not just a single ad hoc patch, for normalized EL ontologies (Theorem 1).","The framework works with a training-free embedding model (SBERT) as well as a fine-tuned ontology embedding (OnT), with the fine-tuned version handling complex existential-restriction bridges better.","On hard negative samples (plausible-looking but false subsumptions), NeurOWL improves F1 by up to 0.13 over the strongest baseline in the reported experiments."],"fun_headline_variants":["LLM-guided search finds missing axioms for subsumptions","NeurOWL: prove plausible is-a statements with missing axioms","Bridge concepts unlock missing axioms in incomplete ontologies","End-to-end neural-symbolic reasoning for OWL ontologies"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The recovery guarantee in Theorem 1 assumes that NeurOWL's neural components actually identify the correct missing axioms at each step; the theorem's conditions only ensure the candidate concept appears in the Stage 3 set, not that it survives top-k retrieval and the LLM's YES/NO test.","fun_headline_variants_meta":{"raw":{"variants":["LLM-guided search finds missing axioms for subsumptions","NeurOWL: prove plausible is-a statements with missing axioms","Bridge concepts unlock missing axioms in incomplete ontologies","End-to-end neural-symbolic reasoning for OWL ontologies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1122,"prompt_tokens":734,"completion_tokens":388,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":319}},"tokens_in":478,"tokens_out":388,"duration_ms":4844,"temperature":1.0,"reasoning_tokens":319,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T01:52:31.869566+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Feed NeurOWL a normalized EL ontology with a known minimal solution S, remove one axiom, and query the resulting non-entailed subsumption. The theorem predicts iterative NeurOWL returns exactly S; if repeated runs on such constructed cases return a different or incomplete set while the correct bridge is in the candidate space, the recovery claim is false. A sharper test: pick a case where the ground-truth bridge is in the top-k but the LLM rejects it while accepting a plausible wrong bridge; systematic occurrence of this pattern would falsify the explanation-accuracy claims even with perfect r","supporting_citations":[],"review_version":2}