{"id":"66d1bbb6-c701-4e48-93e5-f42b228eed82","arxiv_id":"2507.03697","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TUNSR unifies propositional and first-order logic reasoning in one model for knowledge graph link prediction and reports state-of-the-art results on 19 datasets across four reasoning scenarios.","lead":"Researchers propose TUNSR, a single framework that combines neural and symbolic reasoning for knowledge graph link prediction across four settings: transductive, inductive, interpolation, and extrapolation. The model builds a query-specific reasoning graph, passes messages with attention, and extracts human-readable first-order logic rules, reporting gains over many baselines on 19 datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (13)'s '≈' is the load-bearing link between TUNSR's attention scores and FOL rule learning; it is unproven and the RHS as written counts disconnected node tuples, so the FARI rules may not reflect the model's actual scoring.","rationale":"The reader's weakest assumption correctly identifies Eq. (13) as the hinge of the paper's neurosymbolic claim. The paper needs the step-wise attention computation to be a faithful surrogate for summing confidences over grounded FOL rules; otherwise the FARI-induced rules are not an explanation of the model's predictions. I agree that this is the most load-bearing concern, and I strengthen it by noting that the RHS of Eq. (13), taken literally, does not even enforce path connectivity between the nodes summed at successive layers. The inductive scenario's use of FOL-only reasoning is a genuine limitation, but it does not threaten the core claim as directly: the model could still be 'unified' across scenarios if the FOL mechanism were genuinely rule-based. The absence of code and data is important for reproducibility but is not a correctness objection. Because the concern is about an unproven but plausible approximation, the appropriate verdict remains CONDITIONAL: the central claim should not be rejected outright, but it should not be accepted until the equivalence in Eq. (13) is either proven for the model's actual computation or replaced by an explicit, verified path-based formulation.","tokens_in":23671,"tokens_out":5359,"duration_ms":68556,"concrete_test":"On a small KG (e.g., a WN18RR subset), run a trained TUNSR with L=2 and compute the exact sum-over-all-grounded-rules score from Eq. (13)(a) using the rules extracted by FARI (Alg. 1) with their assigned confidences; compare this with the model's actual FOL attention score β_o from Eq. (12) for every query. If the rank correlation or absolute difference exceeds a small tolerance, the '≈' in Eq. (13) is not a faithful description and the claimed FOL-rule reasoning does not match the model's behavior. Additionally, analytically expand Eq. (13) for L=2 and check whether the RHS can be rewritten as a sum over connected paths only; if not, the approximation is not merely unproven but structurally invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TUNSR performs interpretable FOL reasoning rests on Eq. (13), where the sum over grounded FOL rule confidences (part a) is identified, via an unproven '≈', with a step-wise product of sums over reasoning-graph nodes (part b). This is not a harmless shorthand: as written, part (b) multiplies independent sums over all nodes at each layer, which includes combinations of nodes that are not connected by edges, whereas part (a) sums only over actual grounded rule bodies. The subsequent FARI algorithm (Alg. 1) extracts individual paths and assigns confidences by multiplying FOL attentions, but nothing guarantees that the sum of these extracted rule confidences reproduces the model's actual score in Eq. (15), which also mixes propositional embeddings x_o, FOL embeddings y_o, and the learned weight λ. Thus the claimed FOL-rule semantics and interpretability are not derived from the model; they are a post-hoc labeling of attentions. This is load-bearing because without Eq. (13), TUNSR is a neural path-ranking model with rule-looking outputs, and the 'unified neurosymbolic' claim is unsupported. The inductive-scenario limitation (Section 4.3 uses FOL only) is real but secondary: the methodological unity is already at risk if the FOL-attention mechanism is not actually performing FOL reasoning.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TUNSR, a graph-based neurosymbolic framework for knowledge graph link prediction that targets transductive, inductive, interpolation, and extrapolation reasoning. It constructs a query-specific expanding reasoning graph, runs forward message passing that tracks both propositional (entity-dependent) embeddings/attentions and first-order-logic (entity-independent) embeddings/attentions, and introduces a Forward Attentive Rule Induction (FARI) algorithm that extracts Horn rules from the learned FOL attentions. The paper reports state-of-the-art or competitive results on 19 datasets, with ablations, hyperparameter studies, and case studies. The central conceptual claim is that the attention-based computation is equivalent to differentiable merging of grounded FOL rules, which is formalized by the approximation in Eq. (13).","tokens_in":23947,"tokens_out":7785,"duration_ms":86381,"significance":"If the theoretical link were established, TUNSR would be a significant contribution: one architecture spanning four reasoning scenarios with rule-level interpretability and consistently strong empirical performance. The experimental effort is substantial and includes many baselines, 19 datasets, ablation and hyperparameter analyses, and concrete rule examples, and the authors are transparent about the inductive-scenario limitation. The significance is contingent, however, because the equivalence in Eq. (13) and the fidelity of FARI's extracted rules to the model's actual scoring function are not demonstrated; without them, the neurosymbolic interpretation is a post-hoc label rather than a derived property of the model.","major_comments":[{"comment":"The approximation \\(S_o = \\sum_{\\gamma\\in\\Gamma}\\beta_\\gamma \\approx \\prod_{l=1}^L \\sum_{n_j\\in O_l} \\bar f_l(\\mathrm{GRU}(g_{\\bar r}, o_l^{n_j}))\\) is load-bearing for the claim that TUNSR performs FOL reasoning, but it is asserted without derivation. On the left, the sum is over grounded rules whose bodies are connected paths ending at the target entity \\(o\\); on the right, the product of independent sums over all nodes in each layer includes combinations of nodes that are not connected by edges in the reasoning graph, and the expression does not restrict the final-layer node to be \\(o\\). This is not a harmless shorthand, and the paper needs either a proof with the appropriate connectivity constraints or a careful statement of the additional assumptions under which the approximation holds.","section":"Section 3.3, Eq. (13)"},{"comment":"FARI multiplies FOL attention values along paths to compute rule confidences, but the final prediction score in Eq. (15) combines propositional and FOL terms through the learned weight \\(\\lambda\\) and a linear layer \\(W_5\\). The manuscript never shows that the sum of FARI's extracted rule confidences equals or approximates the score used for ranking, so the induced rules in Tables 7-8 are not demonstrably the model's reasoning mechanism. A fidelity experiment, such as ranking candidates using only FARI-derived rule scores and comparing with TUNSR's full rankings, would directly address this gap.","section":"Section 3.3, Algorithm 1, Eq. (15)"},{"comment":"The text states 'As inductive reasoning is entity-independent, we only conduct experiments using FOL reasoning for it.' This admitted limitation conflicts with the paper's headline claim of a unified propositional-and-FOL model across all four reasoning scenarios. The authors should clarify whether the inductive results in Tables 3-4 use the full TUNSR or the FOL-only configuration, and should ideally include the propositional and unified variants in the inductive ablation as well.","section":"Section 4.3"}],"minor_comments":[{"comment":"The symbol \\(o_l^{n_j}\\) in part (b) is not defined; it should presumably be the FOL embedding \\(y_{n_j}^l\\) of the node.","section":"Eq. (13)"},{"comment":"The text refers to 'DURM [32]' but the cited work is DRUM; please correct the typo.","section":"Section 3.3"},{"comment":"The caption says the table includes ICEWS14, ICEWS0515, and ICEWS18, but only ICEWS14 and ICEWS0515 columns appear; please reconcile the caption with the table.","section":"Table 5"},{"comment":"The phrase '14,17% Hits@10 improvement' contains a comma where a decimal point is intended; please correct.","section":"Section 4.2"},{"comment":"The algorithm's notation \\(e^2_{i,l-1}\\) and \\(e^2_{n^{l-1}_i,\\bar r,n^l_j}\\) is not connected to the \\(\\beta\\) attention variables defined in Eqs. (10)-(12), making the rule-confidence update hard to follow.","section":"Algorithm 1"},{"comment":"The model name is rendered inconsistently as 'TUNSR', 'T UNSR ', and 'Tunsr'; please standardize the spelling.","section":"Abstract and Section 1"}],"recommendation":"major_revision","confidential_remarks":"The empirical contribution is strong, but the theoretical framing is substantially ahead of what is proven. I see a well-defined path to a publishable revision: either provide a rigorous derivation of Eq. (13) under explicit connectivity assumptions, or add an empirical fidelity check showing that FARI-induced rules reproduce TUNSR's rankings. Without one of these, the neurosymbolic interpretation is not established. This is why I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: TUNSR is the paper you'd cite if you want one model that works across transductive, inductive, interpolation, and extrapolation KG link prediction. It extends the authors' TECHS by splitting propositional and FOL streams and adding FARI rule extraction. The experiments are the real contribution: 19 datasets, SOTA or near-SOTA against a wide baseline set, ablations show the unified model beats either stream alone, and the case studies show the induced rules are plausible. That is a serious empirical package.\n\nWhat's actually new: the separation of entity-dependent propositional and entity-independent FOL streams in one forward pass, and the explicit application of the same reasoning-graph construction to four scenarios. The components are borrowed—GRU, attention, reasoning graphs, rule induction from Neural LP/DRUM/RED-GNN—but the combination over four scenarios is new.\n\nThe soft spots are real but narrower than the stress-test note suggests. Eq. (13) is indeed handwavy: the RHS as written multiplies independent sums over nodes and would include disconnected combinations, and the '≈' is doing a lot of work. But the actual computation in Eqs. (9)–(12) is a path-probability propagation on the reasoning graph: FOL attention at a node is the accumulated mass over relation paths, and FARI multiplies edge attentions along actual paths. So the path-products do reconstruct the FOL attention score; the paper just never states that properly and instead gestures at an unproven approximation. That's a presentation and formalization flaw, not necessarily a broken core. What is unsupported is the stronger claim that TUNSR is 'doing FOL reasoning' in the symbolic sense. The FOL stream is a differentiable relation-path ranker, and the induced rules are post-hoc labels for that ranker, not rules used to derive predictions. Same criticism applies to Neural LP; worth saying plainly.\n\nThe inductive scenario is a genuine limitation the paper admits in Section 4.3: it runs FOL-only there, so the 'unified' claim doesn't fully cover that setting. Also no code or data. That matters for reproducibility, but the datasets are standard and the hyperparameters are described well enough that reproduction is plausible.\n\nFor a referee: I'd send it out. The empirical breadth alone merits serious review, and the theoretical framing is fixable. The referee report should focus on whether the authors can either prove the connection between the attention aggregation and FARI, or reframe the claims as 'differentiable relation-path ranking with rule outputs.'","headline":"A serious empirical unification of four KG reasoning scenarios, with the neurosymbolic framing stronger than the formal support; worth refereeing, but the authors need to fix Eq. (13) and tone down what 'FOL reasoning' means.","tokens_in":24483,"tokens_out":3148,"would_cite":true,"duration_ms":40618,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TUNSR is a unified neurosymbolic framework that performs both propositional and first-order logic reasoning for knowledge-graph link prediction, reporting state-of-the-art or competitive results across 19 datasets in four reasoning…","keywords":["knowledge graph reasoning","neurosymbolic reasoning","link prediction","first-order logic rules","propositional reasoning","temporal knowledge graphs","rule induction","interpretability"],"falsifier":"On a small knowledge graph, train TUNSR, then for a set of queries enumerate every grounded FOL rule that the FARI algorithm can induce, compute the exact confidence sum from part (a) of Eq. (13), and compare it with the step-wise score from part (b); a large or systematic gap would show that the induced rules are not doing the scoring work the paper attributes to them. A second check is to ablate the FOL stream and ask whether the highest-confidence induced rules still predict the same answers, which would separate genuine rule usage from post hoc explanation.","tokens_in":23447,"feed_emoji":"🧠","tokens_out":9970,"duration_ms":110046,"temperature":0.7,"pith_summary":"TUNSR is a single neurosymbolic architecture for knowledge-graph link prediction that claims to combine propositional reasoning (entity-dependent multi-hop path scoring) with first-order logic reasoning (entity-independent rule application) in one differentiable model. The paper argues that both reasoning styles can be expressed over a uniformly constructed \"reasoning graph\" grown from the query entity, so that one network can serve transductive, inductive, interpolation, and extrapolation settings without architectural changes. It further claims that first-order rule confidence can be computed by merging possible relations step by step rather than by enumerating every rule grounding, and that the resulting attention weights can be read back into explicit, human-readable logic rules. The reported results on 19 datasets across the four scenarios, together with case studies and ablations, are offered as evidence that the unified mechanism is accurate as well as interpretable.","feed_headline":"Unified neurosymbolic model tops 19 KG reasoning datasets","feed_subtitle":"It merges entity-dependent paths with first-order logic rules to give readable explanations while matching or beating specialized baselines.","key_machinery":"The load-bearing object is the reasoning graph, a tree rooted at the query entity (or query entity-time pair for temporal graphs) that expands by searching posterior neighbors and carries a self-loop relation so a node can persist across iterations. Over this graph, a forward logic message-passing mechanism runs two parallel streams: propositional message passing computes entity-dependent node embeddings and attention scores via sigmoid and softmax operations, while FOL message passing uses a GRU to fuse the current relation with the preceding node's FOL embedding and computes entity-independent attention. The central identity is Eq. (13), the asserted approximation that the score obtained by aggregating confidences of all grounded FOL rules equals the score obtained by step-wise, attention-based merging of possible relations; this is what lets rule confidence be accumulated online. The FARI algorithm closes the loop by multiplying FOL attentions along graph paths, adding relations to rule bodies, and deleting self-loop atoms to enumerate all possible Horn rules up to length L.","core_discovery":"On its own terms, the central discovery is that a single network can carry out both propositional and first-order logic reasoning for knowledge-graph link prediction if the two streams are defined over one expanding reasoning graph. The propositional stream updates node embeddings and attentions from entity-dependent messages, while the FOL stream passes GRU-composed relation embeddings from node to node and accumulates entity-independent attention that acts as rule confidence. The paper's key claimed equivalence is that summing the confidences of all grounded FOL rules (part (a) of Eq. (13)) can be approximated by a step-wise process that merges possible relations at each hop (part (b)), making rule learning and grounding differentiable. The FARI algorithm then induces explicit Horn rules — connected and closed rules for static graphs, plus temporal interpolation and extrapolation variants — from the learned attentions, so predictions come with a readable chain of rules. The experiments across 19 datasets and four reasoning scenarios are presented as evidence for both the accuracy and the interpretability of this unified design.","pith_inferences":["If the Eq. (13) approximation holds beyond these benchmarks, the same \"merge relations instead of enumerating groundings\" recipe could be applied to any structured scoring task that is a sum over grounded rules, such as multi-hop question answering or program synthesis over graphs.","The separate propositional and FOL attention maps enable a diagnostic not reported in the paper: on a failure case, one can check whether the error came from missing path evidence (propositional attention near the correct entity but low FOL confidence) or from over-general rules (high FOL confidence but wrong entity), which would turn interpretability into a debugging tool.","A human study could test the interpretability claim directly: if humans asked to simulate the model using only the induced rules cannot predict its answers, the rules would be post hoc, an outcome the paper's case studies do not rule out.","The learned mixing weight λ in Eq. (14) may itself be a useful dataset statistic; reporting it per scenario could reveal a measurable link between graph density and whether propositional or FOL reasoning dominates."],"forward_implications":["A single TUNSR model answers transductive, inductive, interpolation, and extrapolation queries without changing the architecture; only the node structure of the reasoning graph changes.","The attention weights used for scoring double as rule confidences, so every prediction can be accompanied by the chained FOL rules that produced it.","Because FOL reasoning is entity-independent, the induced rules transfer to unseen entities and unseen graphs that share the relation set, which is what makes inductive reasoning possible.","Combining propositional and FOL streams outperforms either alone across the reported settings, since propositional paths capture entity-specific evidence while FOL rules provide generalization.","Rule length behaves as a bias-variance knob: performance improves up to roughly 6 hops on static and 4 hops on temporal graphs, then stabilizes or degrades as longer rules add noise."],"supporting_citations":[{"why":"It supplies the differentiable rule-learning baseline whose grounding-sum scoring TUNSR re-expresses as step-wise merging.","marker":"[21]"},{"why":"It provides the end-to-end differentiable rule-mining baseline used in transductive and inductive comparisons.","marker":"[32]"},{"why":"It contributes the relational-digraph reasoning graph structure and multi-class log-loss that TUNSR adapts.","marker":"[58]"},{"why":"It is the earlier temporal logical graph network from which this framework generalizes to all four reasoning scenarios.","marker":"[30]"},{"why":"It is the temporal logical-rule baseline for extrapolation comparison and for contrasting rule-induction strategies.","marker":"[28]"},{"why":"It offers the neural rule-learning approach whose rule-grounding aggregation is compared with the differentiable merging.","marker":"[33]"},{"why":"It is the recursive rule-learning baseline used in the transductive comparison.","marker":"[23]"},{"why":"It is the tensor-decomposition temporal baseline for interpolation and extrapolation comparisons.","marker":"[77]"},{"why":"It is the strongest interpolation baseline that TUNSR compares against.","marker":"[27]"}],"fun_headline_variants":["Neurosymbolic KG reasoning: one model, 19 datasets","Unified model beats specialists across 19 KG tasks","A single framework for neural and symbolic KG reasoning","Tunsr merges embeddings and FOL for KG link prediction","One network learns both neural paths and readable rules"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the step-wise, attention-based merging of possible relations in part (b) of Eq. (13) faithfully approximates summing the confidences of every grounded first-order logic rule in part (a); the paper marks this as an approximation without a proof, and if the equivalence fails, the induced rules and attention scores would not be the true drivers of the model's predictions.","fun_headline_variants_meta":{"raw":{"variants":["Neurosymbolic KG reasoning: one model, 19 datasets","Unified model beats specialists across 19 KG tasks","A single framework for neural and symbolic KG reasoning","Tunsr merges embeddings and FOL for KG link prediction","One network learns both neural paths and readable rules"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1473,"prompt_tokens":1030,"completion_tokens":443,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":363}},"tokens_in":646,"tokens_out":443,"duration_ms":5381,"temperature":1.0,"reasoning_tokens":363,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:03:51.840986+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small knowledge graph, train TUNSR, then for a set of queries enumerate every grounded FOL rule that the FARI algorithm can induce, compute the exact confidence sum from part (a) of Eq. (13), and compare it with the step-wise score from part (b); a large or systematic gap would show that the induced rules are not doing the scoring work the paper attributes to them. A second check is to ablate the FOL stream and ask whether the highest-confidence induced rules still predict the same answers, which would separate genuine rule usage from post hoc explanation.","supporting_citations":[{"cited_title":"Yang et al","cited_arxiv_id":null,"evidence_quote":"It supplies the differentiable rule-learning baseline whose grounding-sum scoring TUNSR re-expresses as step-wise merging."},{"cited_title":"Sadeghian et al","cited_arxiv_id":null,"evidence_quote":"It provides the end-to-end differentiable rule-mining baseline used in transductive and inductive comparisons."},{"cited_title":"Zhang and Q","cited_arxiv_id":null,"evidence_quote":"It contributes the relational-digraph reasoning graph structure and multi-class log-loss that TUNSR adapts."},{"cited_title":"Lin et al","cited_arxiv_id":null,"evidence_quote":"It is the earlier temporal logical graph network from which this framework generalizes to all four reasoning scenarios."},{"cited_title":"Liu et al","cited_arxiv_id":null,"evidence_quote":"It is the temporal logical-rule baseline for extrapolation comparison and for contrasting rule-induction strategies."},{"cited_title":"Qu et al","cited_arxiv_id":null,"evidence_quote":"It offers the neural rule-learning approach whose rule-grounding aggregation is compared with the differentiable merging."},{"cited_title":"Cheng et al","cited_arxiv_id":null,"evidence_quote":"It is the recursive rule-learning baseline used in the transductive comparison."},{"cited_title":"Lacroix et al","cited_arxiv_id":null,"evidence_quote":"It is the tensor-decomposition temporal baseline for interpolation and extrapolation comparisons."},{"cited_title":"Li et al","cited_arxiv_id":null,"evidence_quote":"It is the strongest interpolation baseline that TUNSR compares against."}],"review_version":1}