NeurIPS 2026

NEURIPS 2026 / POSTER

Neural Structural ReasonerA Brain-inspired Architecture for
Reasoning over Structured Knowledge

* Equal contribution · † Corresponding author

Author affiliations

1 Beijing Institute for Brain Research, CAMS & PUMC

2 Chinese Institute for Brain Research, Beijing

3 Beijing Key Laboratory of Brain Science and Brain-Machine Interface

4 Sun Yat-sen University

KNOWLEDGE IN CONNECTIVITYREASONING THROUGH DYNAMICS
NSR overview: a relation triple is encoded as connections between entity–relation bindings; four neural populations support learning and query-driven reasoning.
NSR at a glance. Four coupled populations preserve entity and relation identities. Facts are written into R–Z and Z–Z connectivity; query-driven dynamics read out candidate entities. This schematic single-fact encoding and readout shows selected connections, not measured activations.
01

Abstract

Structural reasoning, the ability to recognize and make inferences over the relational structure between objects and concepts, is a hallmark of human cognition, yet prevailing methods often collapse relational topology into flat embeddings, cannot discover hidden structure and lack interpretability. We introduce Neural Structural Reasoner (NSR), a brain-inspired network that preserves relational structure directly in the connectivity and dynamics of coupled neuronal populations. NSR draws inspiration from three biological mechanisms: multi-layered architecture for encoding hierarchical knowledge, stable representations of entity and concepts, and path integration for input-driven state inference. At query time, NSR parallelizes computation over candidate relational structures and leverages confidence-weighted scores to perform link prediction. Across standard knowledge-graph benchmarks, NSR achieves competitive accuracy without leading on every dataset, and has lower reported training times than several neural baselines. Because reasoning is implemented through sequences of human-readable neuron activations, NSR affords native interpretability by tracking intermediate inference steps. The model further extracts latent relational hierarchies and compositional rules, demonstrating the brain-inspired architecture as an effective, efficient, and highly interpretable substrate for structural reasoning.

01 / REPRESENTATION

Explicit relational structure

Entities, bindings, relations and compositions share one connected architecture.

02 / LEARNING

Reusable associations

Local association learns relation correspondences and compositional structure.

03 / INFERENCE

Inspectable reasoning

Intermediate entities and relations remain visible in the network dynamics.

02

Model architecture

Paper §3 ↗

NSR separates entities, directed relations, and relation compositions into four coupled populations. The binding population Z associates each entity with its relational roles; connectivity preserves the structure shared across facts and queries.

Brain regions inspiring NSR: PFC, ATL, EC and HPC, with their roles in integration, semantic associations, relational maps and relational memory.
Biological inspiration. These correspondences motivate the architecture; they do not assert biological equivalence.
EEntitiesHippocampus · HPC
A dedicated neuron preserves each entity’s identity across facts.
ZEntity–relation bindingsEntorhinal cortex · EC
Relation-indexed binding neurons connect entities to their roles.
RDirected relationsAnterior temporal lobe · ATL
Relations and their inverses support retrieval of learned associations.
CRelation compositionsPrefrontal cortex · PFC
Learned compositions make multi-step relational patterns reusable.
INTERACTIVE METHOD FIGURE

Encoding a relation triple

Edit (h, r, t), or select a stage to inspect the fixed and fact-encoded connections.

TRY A FACTor edit the triple above
SCHEMATIC · EQ. 3
Select a neuron to inspect its role.
Fixed E–Z connectionsFact-encoded connections · weight 1
01 / 04

Keep the entities.

The head and tail each have an entity neuron. E–Z connections are already fixed: each entity connects to its own relation-indexed binding neurons.

THE SAME FACT, IN REVERSE

r⁻¹ is the inverse direction of r.

CCompositions are learned across facts. Single-triple encoding leaves C unchanged.

Illustrative encoding, following §3.2 / Eq. 3. Highlights explain roles; they are not recorded activations. Encoding assigns weights directly—the animation reveals the construction in steps. Read the formulation ↗

03

Learning and interpretable reasoning

Learning. Hebbian-style association learns correspondences between relations and weights for relation compositions. For static graphs, an accelerated offline procedure computes composition weights from encoded connectivity, without simulating each learning episode.

Inference. A query retrieves associated relations or compositions. Entity-conditioned activity traverses the encoded network to reach candidate tails. Grounded-path support is aggregated using a validation-selected max or sum rule to rank candidates; these scores are not calibrated probabilities.

Query Brazil, embassy, unknown: relation activity retrieves weightedunvote inverse; the Brazil entity activates bindings, activity propagates through Z, and the UK entity is read out. The figure shows the logical path and neuronal activations.
Figure 1. Traceable reasoning. The query (Brazil, embassy, ?) retrieves an associated inverse relation, activates entity–relation bindings, and reaches the tail entity UK. The logical path and intermediate neural activations can both be inspected. Adapted from Figure 2 of the paper.
Explore recorded network dynamicsReplay five recorded states, inspect neurons, and compare three Nations queries in the Blog version.
Learned Nations relation compositions involving exportbooks, releconomicaid, relexportbooks, economicaid and embassy.
Figure 2. Learned compositions. Relation chains become reusable structures for multi-step reasoning. Examples learned from Nations.
Projection of learned Nations representations, with concept-like clusters associated with historical political blocs.
Figure 3. Relational organization. Learned representations organize Nations entities into concept-like groups associated with historical political blocs.
04

Accuracy and training efficiency

We evaluate filtered link prediction on Nations, Kinship, YAGO3-10, and FB15k-237. NSR combines competitive, dataset-dependent accuracy with lower reported training times than several neural baselines. The comparison below presents predictive accuracy and training time together; select a dataset or method to inspect the trade-off.

ACCURACY × REPORTED TRAINING TIMEClick a method to compare

Source: paper Tables 1–2 and author slides 8–9. NSR training times refer to the accelerated implementation for static knowledge graphs (Appendix A.2.1). Reported hardware and protocols differ; ratios are indicative, not hardware-controlled speedups. The horizontal axis uses a logarithmic scale. Models without reported times remain in the table. * denotes a baseline value quoted from prior work or official releases.

Full results for the selected dataset
Evaluation scope and limitations

NSR does not achieve the highest accuracy on these benchmarks, and its training time is not the lowest for every dataset or comparison. Evaluation uses discrete, symbolically specified knowledge graphs. Generalization to noisy, temporal, or perceptual inputs remains open; longer relation chains also pose scaling challenges. See the paper for experimental protocols, implementation details and limitations.

05

Author presentation

Download slides ↓
4 min 28 sec · English · SubtitledDownload video ↓
06

Citation

@misc{jia2026neuralstructuralreasoner,
  title={Neural Structural Reasoner: A Brain-inspired Architecture
         for Reasoning over Structured Knowledge},
  author={Zixing Jia and Yuhang Pan and Ni Ji},
  year={2026},
  eprint={2609.36620},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2609.36620}
}