Neural Structural Reasoner: A Brain-inspired Architecture for Reasoning over Structured Knowledge

Motivation

• Structural reasoning concerns inferences over relations between objects and concepts, which humans use to reason beyond entities alone.Existing methods tend to flatten rich structure and can obscure how answers were reached.Source: §1 · pp. 1–2
• Existing biologically inspired models draw on neural principles, but the authors identify a lack of modular brain-inspired architecture for encoding and reasoning over structured knowledge at scale.Source: §1 · p. 2

Contributions

• NSR represents relational structure through network connectivity and dynamics, reasons in parallel over candidate structures, ranks answers with confidence scores, and exposes human-readable inference paths.Its reported accuracy is competitive but dataset-dependent, with relatively less training time than several neural baselines.Source: §1 · p. 2

2.1 Structural reasoning across machine learning

• Knowledge-graph link prediction makes structural reasoning concrete: score candidate tails for a query (h, r, ?) using the graph's typed relations.Existing approaches include embeddings, message-passing networks, rule mining, and language models, with limitations in explicit composition, traceability, or robustness to missing and noisy edges.Source: §2.1 · p. 2

2.2 Neural substrates of structural reasoning

• The reviewed brain systems support stable concepts, relation-specific binding, and compositional generalization; path integration may update internal state through accumulated transitions.The passage notes that direct evidence for extending path-integration dynamics to abstract reasoning remains scarce.Source: §2.2 · p. 3

2.3 Biologically inspired models of structural reasoning

• Prior biologically inspired models offer ingredients for relational cognition, but the review identifies a gap in biologically plausible, input-driven transitions across high-dimensional relation spaces.The cited models are limited respectively by backpropagation-trained dynamics, unexplored abstract reasoning, low-dimensional navigation, or 2D-torus dynamics.Source: §2.3 · p. 3

3.1 Model Architecture

• NSR uses four layers to encode entities, relation-driven state transitions, relations, and relation compositions within a hierarchical network.Source: §3.1 · p. 3

3.1.1 Network Connectivity and Dynamics

• The layers form a bidirectional hierarchy, with fixed entity couplings, learnable remaining weights, and gated dynamics that allow intermediate entities and relations to be read from activations.Figure 1 distinguishes fixed connections, observed-triple connections, and Hebbian-learned connections.Source: §3.1.1 · pp. 3–4, Figure 1 · p. 4

3.2 Encoding of Relation Triples

• Relation and inverse-relation neurons are connected with entity-encoding neurons when relation triples are encoded.Source: §3.2 · p. 4

3.2.1 Reasoning over (hi, rk, ?) through Path-Integration-Inspired Dynamics

• For a query (h_i, r_k, ?), LZ dynamics activate relation-conditioned states and then a tail-entity neuron, following a path-integration-inspired computation.The sequence is inspired by path-integration models proposed for entorhinal grid cells.Source: §3.2.1 · p. 5

3.3 Learning Compositional Rules among Relations

• Hebbian-like learning registers compositional rules when a two-hop chain closes a loop to the original head, linking the chain to the corresponding relation.For same-head queries with matching tails, co-activated relation neurons are coupled; the dynamics also extend to symmetric and inverse relations.Source: §3.3 · p. 5

3.4 Reasoning Phase

• NSR retrieves equivalent relations and compositions for the query relation, reasons over each, and retains grounded paths with their candidate tails and support scores.Equivalent-relation activations and composition-neuron activations provide path-support scores for ranking candidates.Source: §3.4 · p. 6
• Candidate scores aggregate support across grounded paths by maximum or sum, with the aggregation mode selected on validation data.The scores rank candidates rather than representing calibrated probabilities; maximum aggregation may overvalue a spurious path, while summation may overcount correlated paths.Source: §3.4 · p. 6

4.1 Experimental setup

• The evaluation compares filtered MRR and Hits@1/3 plus training time on Nations, Kinship, YAGO3-10, and FB15k-237 against ten baselines.The main tables cover Nations and Kinship, and YAGO3-10 and FB15k-237, respectively.Source: §4.1 · p. 6, §4.2 · p. 7, Table 1 · p. 7

4.2 Empirical performance

• Across four benchmarks, NSR ranks near the top on Nations and YAGO3-10, remains competitive on FB15k-237, and is weaker on Kinship while training faster than several neural baselines.The reported performance is dataset-dependent rather than uniformly superior; timings are indicative because some are published or cross-hardware comparisons.Source: §4.2 · pp. 7–8, Table 1 · p. 7, Table 2 · p. 7

4.3 Traceable reasoning steps

• For (Brazil, embassy, ?), embassy activates weightedunvote−1 as the dominant equivalent relation, and recurrent LZ dynamics lead to the UK tail activation.The human-readable activation sequence makes intermediate inference steps inspectable and errors auditable by the activated relation or where LZ dynamics diverge.Source: §4.3 · p. 8, Figure 2 · p. 8

4.4 Learning of latent structures

• NSR learns context-sensitive relation compositions: exportbooks + releconomicaid → embassy receives confidence 0.93, while repeating embassy twice does not yield a high-confidence rule.On Kinship1990_EXTENDED, deterministic kinship rules can receive confidence 1 after a single observation by integrating prior logical constraints through hyperparameters.Source: §4.4 · pp. 8–9, Figure 3 · p. 9
• In Nations, recurrent relational connections organize entities into clusters corresponding to Cold War political blocs, including distinct Western, Eastern, and non-aligned groups.The learned structure combines a hierarchy of relation compositions with a topology organizing entities.Source: §4.4 · p. 9, Figure 4 · p. 9

5 Conclusion

• NSR combines hierarchical representation, relation-driven state transitions, and Hebbian-style learning to reason over knowledge graphs.It preserves relational structure in layered connectivity rather than collapsing it into flat embeddings.Source: §5 · pp. 9–10
• NSR achieves competitive, dataset-dependent benchmark accuracy, exposes human-readable reasoning traces, and discovers compositional rules and relational similarity among entities.The paper presents these properties as a promising step toward efficient and interpretable structural-reasoning models.Source: §5 · p. 10
• Evaluation is currently limited to discrete triplet-structured knowledge graphs, leaving generalization to temporal, hyper-relational, event-based, or noisy real-world knowledge open.The model also assumes entities and relations are already symbolically specified; scaling rule discovery to larger graphs and longer chains remains a challenge.Source: §5 · p. 10

A.1 Learning and Encoding of KG Training Data

• Training triples and their inverses set connectivity weights, while co-occurring relations on the same entity pair drive Oja-like updates that capture equivalence, symmetry, and inverse structure.Source: §A.1 · p. 16

A.2 Compositional Rules Learning

• Sequence-selective neurons detect ordered relation pairs through asymmetric-delay coincidence detection, and concurrent activation strengthens their synapses to consequent relations.Source: §A.2 · p. 16

A.2.1 Offline learning on static graphs

• On static graphs, sparse adjacency products enumerate two-hop experiences, allowing a single-pass closed-form computation of the same compositional weights as online learning.The full dynamical and offline variants attain comparable performance on all benchmarks.Source: §A.2.1 · p. 17

B.1 Reasoning phase (algorithmic form)

• The reasoning algorithm retrieves equivalent relations and compositions, grounds one- and multi-hop paths, assigns support scores, and aggregates scores for each candidate tail.Source: §B.1 · p. 18

• T_thresh admits a relation to the equivalent-relation set only when its activation meets the threshold; for Countries, it is set to infinity because the dataset has only two relations.Source: §B.2 · p. 18
• The simplified implementation uses one scalar δθ for both relation and composition learning rates because it counts in one pass rather than using iterative Oja updates.Source: §B.2 · p. 18
• τs gates composition neurons by chain evidence and τc gates consequent-relation readout by weight; Optuna selected the settings, to which NSR is not highly sensitive.The hyperparameters are grouped by processing stage.Source: §B.2 · p. 18, Table 3 · p. 18

C.1 Datasets

• The evaluation includes standard KG benchmarks from compact relational graphs to larger graphs, plus Kinship1990_EXTENDED, designed to stress-test compositional reasoning.Table 4 summarizes dataset entities, relation types, and triples by split.Source: §C.1 · p. 20, Table 4 · p. 20

C.2 How to Construct Kinship1990_EXTENDED

• Kinship1990_EXTENDED addresses the original dataset’s lack of symmetric relations, flat labels, and small scale by adding symmetric sibling links and five explicit chain-relation types.The chain labels are abstracted from multi-hop genealogical paths, encouraging reuse of primitive relations rather than memorization of atomic mappings.Source: §C.2 · p. 20
• The extended graph has 480 entities, 14 relation types, and 2,240 triples, with a 70/10/20 train-validation-test split; held-out triples are predominantly from extended relations.Base-relation edges needed for genealogical consistency appear in training, so accurate test prediction benefits from compositional reuse of the trained primitives.Source: §C.2 · p. 21

• Neural embedding baselines use PyKEEN under identical splits, with dimensions of 100 for small datasets and 200 for large datasets, and training capped at 150 epochs.DistMult, ComplEx, RESCAL, and ConvE use LCWA; TransE and RotatE use sLCWA.Source: §C.3 · p. 21, Table 5 · p. 21
• RNNLogic mines relational paths up to length 3 and learns rule weights with a lightweight predictor, without pretrained or learned entity and relation embeddings.Source: §C.3 · p. 21, Table 6 · p. 21
• Baseline and NSR performance is reported across five random seeds, and all methods use identical train, validation, and test splits for each dataset.Source: §C.3 · p. 21

Countries S3

• NSR solves Countries S3 perfectly, where test queries require composing neighborOf and locatedIn chains absent as direct country–continent training facts.The best embedding baseline reaches MRR 0.199 and RNNLogic reaches 0.341; the reported comparison attributes their weaker results to static vector similarity and externally mined templates, respectively.Source: §C.4 · pp. 21–22, Table 7 · p. 22

Kinship1990_EXTENDED

• On Kinship1990_EXTENDED, NSR attains the best MRR (0.953) and Hits@1 (94.3), surpassing all baselines on composition-focused test triples.DistMult is strong on MRR (0.941), while the authors contrast this result with classical Kinship, where ConvE leads.Source: Table 8 · p. 22, §C.4 · p. 22

WN18RR

• On WN18RR, NSR achieves the best Hits@1 (0.463), is competitive with the strongest embedding models on MRR, and remains weaker than NBFNet.The paper characterizes NSR as competitive rather than uniformly superior on this benchmark.Source: §C.4 · p. 22, Table 9 · p. 23

Comparison with symbolic and neuro-symbolic reasoners

• NSR outperforms evaluated differentiable rule learners on every completed dataset, performs comparably to strong symbolic rule miners, and leads on YAGO3-10; NBFNet leads on WN18RR and FB15k-237, while AMIE leads on Nations.The authors note that these heterogeneous large benchmarks are not designed around NSR’s compositional structure, so performance is not confined to small, composition-aligned benchmarks.Source: §C.4 · p. 22, Table 10 · p. 23

Training and inference cost

• NSR trains in seconds to minutes across benchmarks, with modest inference costs, but cross-method timing comparisons are indicative rather than strict speedups.The timing table includes measurements from differing hardware and stopping criteria.Source: §C.4 · p. 22, Table 11 · p. 23

C.5 Ablation Study

• Disabling Hebbian learning causes the largest overall MRR drops (−0.45, −0.60, −0.93), while inverse encoding, relation-equivalence retrieval, and compositional inference have distinct dataset-dependent effects.Inverse encoding is decisive on kinship benchmarks; retrieval contributes across all datasets, and composition matters most on Kinship1990_EXTENDED. With learned matrices at zero, residual ranks are effectively determined by tie-breaking noise.Source: §C.5 · p. 24, Table 12 · p. 24
• These ablations test the computational roles of implemented components, not the biological realism of the mechanisms that inspired them.The paper presents those mechanisms as computational inspiration.Source: §C.5 · p. 24
• Max aggregation can favor a single spurious path, whereas sum aggregation can overcount correlated paths, so the aggregation mode is selected on validation data.For one Kinship query, sum aggregation ranks the target first after pooling support from 132 grounded paths, unlike max aggregation.Source: §C.5 · pp. 24–25

D.1 Symmetry Reasoning Trace

• Symmetry reasoning retains the original relation neuron while an isomorphic inverse branch activates the inverse-relation neuron; Figure 5 displays only the symmetry branch.Both branches use the same gated-diffusion and binding structure.Source: Figure 5 · p. 25, §D.1 · p. 25

D.2 Latent Relational Structure of Countries

• The Countries dataset has country, region, and continent strata, but observed neighborOf edges connect only countries, making higher-level adjacency a probe of generalization across the hierarchy.The dataset’s two primitive relations are neighborOf and locatedIn.Source: §D.2 · p. 25
• Two chained rules lift neighborOf from countries to regions and then from regions to continents using locatedIn links.The rules use nb for neighborOf and loc for locatedIn.Source: §D.2 · p. 25
• Without any region- or continent-level neighbor edges in training, NSR recovers 29 region-neighbor and 4 continent-neighbor edges that are geographically coherent.The inferred continent adjacencies include the Africa–Europe–Asia triangle and an Asia–Oceania link; Figure 6 visualizes the layered graph.Source: Figure 6 · p. 26, §D.2 · p. 26