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

Zixing Jia; Yuhang Pan; Ni Ji

cs.AISep 29, 2026

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

Authors: Zixing Jia, Yuhang Pan, Ni Ji

Organizations: Beijing Institute for Brain Research, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, 102206, China · Chinese Institute for Brain Research, Beijing; Beijing, 102206, China · Beijing Key Laboratory of Brain Science and Brain-Machine Interface · Sun Yat-sen University, Guangzhou, China

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.

Figures & tables

Figure 1: The Neural Structural Reasoner Architecture. (a) Four core layers of NSR: $L_{E}$ encode entities; $L_{Z}$ instantiates relation-driven state transition; $L_{R}$ and $L_{C}$ encode relations and their compositions (e.g., chains). Connections in black are fixed a priori, those in green encode observed triples, those in orange are learned via Hebbian updates to capture latent relational structure. (b) Sequence of neuron activation that implements one-step reasoning over $(h,r,\,?)\mapsto t$ .

		Nations			Kinship			Training time
Category	Method	MRR	H@1	H@3	MRR	H@1	H@3	Nations	Kinship
Embedding- based	ConvE	0.8029	69.05	89.45	0.7927	68.11	88.45	26 s	64 s
Embedding- based	RotatE	0.5351	33.23	66.02	0.7598	63.37	86.26	24 s	84 s
Symbolic rule learning	AnyBURL	0.7994	69.15	89.55	0.6768	54.10	76.63	64 s	63 s
Symbolic rule learning	AMIE	0.8559	77.11	92.54	0.6767	55.03	76.26	0.6 h	2 s
Neural rule learning	NeuralLP	0.6841	52.74	81.59	0.6072	47.30	68.06	40 s	26 s

Table 1: Link-prediction performance on Nations and Kinship. Hits@1 and Hits@3 are percentages. Higher ranking scores and lower training times are better; the best value in each column is bold.

		YAGO3-10			FB15k-237			Training time
Category	Method	MRR	H@1	H@3	MRR	H@1	H@3	YAGO3-10	FB15k-237
Embedding- based	ConvE	0.6365	59.03	71.28	0.4095	31.61	44.89	10.3 h	991 s
Embedding- based	RotatE	0.1812	9.98	21.04	0.3368	27.07	41.51	1.5 h	1.5 h
Symbolic rule learning	AnyBURL	0.5589	50.78	60.16	0.332 ∗	24.7 ∗	—	1000 s	1000 s ∗
Symbolic rule learning	AMIE	0.5473	49.58	59.27	0.2170	16.57	23.07	83 s	8 s
Neural rule learning	NeuralLP	—	—	—	0.3166	24.45	34.31	—	11.9 h

Table 2: Link-prediction performance on YAGO3-10 and FB15k-237. Hits@1 and Hits@3 are percentages. Higher ranking scores and lower training times are better; the best available value in each column is bold. A dash indicates a missing entry.

Figure 2: Traceable reasoning in NSR. (a) A schematic of the inferred logical path: embassy triggers its inverse to map Brazil to UK . (b) Neuronal activations at each stage of the query (Brazil, embassy, ?) . Activating the relation neuron embassy triggers the activation of candidate equivalent relations (e.g., weightedunvote -1 is an associated inverse relation of embassy ) in the $L_{R}$ layer. The selected relation and head entity Brazil jointly project onto the $L_{Z}$ layer. State transition in $L_{Z}$ through recurrent dynamics leads to activation of neurons encoding the tail entity UK .

Figure 3: Latent relational hierarchy. Compositional rules learned by NSR, satisfying $(h,r_{1},x)\wedge(x,r_{2},t)=(h,r_{3},t)$ , on (a) Nations and (b) Kinship1990_EXTENDED. Solid lines denote $r_{1}$ , dashed lines denote $r_{2}$ , and color intensity indicates rule confidence.

Figure 4: Learned relational organization in Nations. PCA projection of nations based on their relational connections in $L_{Z}$ . Four clusters emerge corresponding to Cold War political blocs.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	T_thresh	$\tau_{s}$	$\tau_{c}$	$\delta_{\theta}$
Nations	0.22	23	0.20	0.52
Kinship	0.90	27	0.17	0.20
Countries S3	$\inf$	1	1.00	1.00
Kinship1990_EXTENDED	0.46	23	0.13	0.34

Appendix

Table 3: Hyperparameters for the main KG benchmarks

Dataset	Entities	Relations	Total	Train	Val	Test
Nations	14	55	1,992	1,592	199	201
Kinship	104	26	10,686	8,544	1,068	1,074
Kinship1990_EXTENDED	480	14	2,240	1,568	224	448
Countries_S3	271	2	1,033	985	24	24
WN18RR	40,943	11	93,003	86,835	3,034	3,134
FB15k-237	14,541	237	310,116	272,115	17,535	20,466

Appendix

Table 4: Statistics of the datasets used in our experiments. All counts in the last five columns refer to the number of edges (triples).

Model	Dim	Loop	Inverse	LR	Epochs	Special
TransE	100	sLCWA	No	$1\mathrm{e}{-3}$	150	scoring_fct_norm=1
DistMult	100	LCWA	No	$1\mathrm{e}{-3}$	150	—
ComplEx	100	LCWA	No	$1\mathrm{e}{-3}$	150	—
RotatE	100	sLCWA	No	$5\mathrm{e}{-4}$	150	—
ConvE	100,200	LCWA	Yes	$1\mathrm{e}{-3}$	150	out_ch=32, dropouts $\{0.2,0.2,0.3\}$
RESCAL	100	LCWA	No	$1\mathrm{e}{-3}$	150	—

Appendix

Table 5: Neural baseline hyperparameters and training configuration.

Model	Rule Len	Hidden Dim	LR	Iters	Special
RNNLogic (Miner)	3	—	—	—	16 threads
RNNLogic (Predictor)	—	32	$5\mathrm{e}{-3}$	10	batch=16, smoothing=0.1, expectation=True, without embedding

Appendix

Table 6: Symbolic baseline hyperparameters and training configuration.

Model	MRR $\uparrow$	Hits@1 $\uparrow$	Hits@3 $\uparrow$	Train Time (s)
ConvE	0.1987 $\pm$ 2.47	11.25 $\pm$ 2.50	19.17 $\pm$ 4.04	20.3 $\pm$ 2.7
DistMult	0.1761 $\pm$ 1.42	7.50 $\pm$ 1.02	20.42 $\pm$ 3.06	17.4 $\pm$ 1.0
RotatE	0.1184 $\pm$ 2.42	4.17 $\pm$ 1.86	10.00 $\pm$ 4.45	19.4 $\pm$ 2.9
RESCAL	0.1906 $\pm$ 1.00	10.42 $\pm$ 2.64	18.33 $\pm$ 2.04	14.7 $\pm$ 2.1
ComplEx	0.0290 $\pm$ 2.68	0.83 $\pm$ 1.67	2.08 $\pm$ 4.17	18.4 $\pm$ 2.2
TransE	0.1182 $\pm$ 0.61	0.00 $\pm$ 0.00	14.17 $\pm$ 2.76	18.9 $\pm$ 1.7

Appendix

Table 7: Performance comparison on Countries S3. Best results in each column are highlighted in bold. Hits@

k

are reported as percentage values in

[0,100]

; standard deviations are in units of

10^{-2}

(MRR) and percentage points (Hits@

k

Model	MRR $\uparrow$	Hits@1 $\uparrow$	Hits@3 $\uparrow$	Train Time (s)
ConvE	0.8265 $\pm$ 0.91	74.93 $\pm$ 1.47	87.83 $\pm$ 0.72	60.5 $\pm$ 7.5
DistMult	0.9409 $\pm$ 0.15	90.62 $\pm$ 0.30	97.75 $\pm$ 0.25	31.9 $\pm$ 4.0
RotatE	0.5988 $\pm$ 4.47	51.94 $\pm$ 5.27	64.40 $\pm$ 4.20	26.0 $\pm$ 1.1
RESCAL	0.0144 $\pm$ 0.28	0.22 $\pm$ 0.12	0.80 $\pm$ 0.31	30.3 $\pm$ 5.3
ComplEx	0.0220 $\pm$ 0.34	0.42 $\pm$ 0.23	1.41 $\pm$ 0.50	31.4 $\pm$ 1.6
TransE	0.2249 $\pm$ 0.58	2.19 $\pm$ 0.69	33.42 $\pm$ 1.58	25.3 $\pm$ 2.4

Appendix

Table 8: Performance comparison on Kinship1990_EXTENDED. Conventions as in Table 7 .

Method	MRR	Hits@1	Hits@3	Hits@10
TransE	.226	—	—	.501
DistMult	.430	.390	.440	.490
ConvE	.430	.400	.440	.520
ComplEx	.440	.410	.460	.510
RotatE	.476	.428	.492	.571
BoxE	.451	.400	.472	.541

Appendix

Table 9: Filtered link prediction on WN18RR. Symbolic and neuro-symbolic baselines report MRR only; “—” indicates unreported or not scalable. Best result in each column in bold.

Method	Nations	Kinship	WN18RR	YAGO3-10	FB15k-237
AnyBURL	0.7994	0.6768	0.5658	0.5589	0.332 ∗
AMIE	0.8559	0.6767	0.4157	0.5473	0.2170
PRA / PathRank	0.5933	0.6296	0.0556	0.4678	0.0972
NTP	0.6223	0.612 ∗	—	—	—
NeuralLP	0.6841	0.6072	0.4677	—	0.3166
NCRL	0.4571	0.6050	0.4070	0.380 ∗	0.300 ∗

Appendix

Table 10: Link prediction results (MRR) on all KG completion benchmarks. Best results are in bold ; asterisks (*) denote values taken from prior work or official repositories, and “—” indicates results not reported or the method does not scale to the dataset.

Method	Nations	Kinship	WN18RR	YAGO3-10	FB15k-237
AnyBURL	64 s / 16 s	63 s / 36 s	603 s / 11 s	1000 s / 424 s	1000 s ∗∗ / —
AMIE	2246 s / $<$ 0.1 s	2 s / 0.3 s	2 s / $<$ 0.1 s	83 s / 0.6 s	8 s / 1.2 s
PRA / PathRank	34 s / 0.5 s	86 s / 0.3 s	22 s / 156 s	239 s / 354 s	229 s / 841 s
NTP	1.9 h / 8 s	—	—	—	—
NeuralLP	40 s / 8 s	26 s / 4 s	1.1 h / 129 s	— (OOM)	11.9 h / 533 s
NCRL	134 s / 43 s	99 s / 7 s	244 s / 24 s	—	—

Appendix

Table 11: Training time / full-test-set inference time (seconds unless noted). ∗∗ quoted from the official AnyBURL-23-1 release; other local runs use the same machine. NBFNet entries use official training profiles, so cross-hardware comparisons are indicative. OOM: out of memory.

Variant	Nations	Kinship	Kinship1990_EXTENDED
Full model	0.81 $\pm$ 0.03	0.65 $\pm$ 0.01	0.95 $\pm$ 0.00
w/o inverse encoding	0.66 $\pm$ 0.03 ( $-$ 0.15)	0.05 $\pm$ 0.00 ( $-$ 0.60)	0.10 $\pm$ 0.00 ( $-$ 0.85)
w/o relation-equivalence retrieval	0.61 $\pm$ 0.01 ( $-$ 0.20)	0.42 $\pm$ 0.00 ( $-$ 0.24)	0.82 $\pm$ 0.00 ( $-$ 0.12)
w/o compositional inference	0.79 $\pm$ 0.03 ( $-$ 0.03)	0.48 $\pm$ 0.01 ( $-$ 0.17)	0.46 $\pm$ 0.00 ( $-$ 0.48)
w/o Hebbian learning	0.36 $\pm$ 0.01 ( $-$ 0.45)	0.05 $\pm$ 0.00 ( $-$ 0.60)	0.02 $\pm$ 0.00 ( $-$ 0.93)

Appendix

Table 12: Component-wise ablations of NSR (MRR, mean

\pm

std over five seeds). Values in parentheses denote

\Delta

MRR relative to the paired full model. Best results in each column are highlighted in bold.

Figure 5: Symmetry-reasoning trace (symmetry branch only). In Path A, symmetry and inverse evaluations are performed in parallel. The complete trace therefore contains two branches: a symmetry branch that retains the original relation neuron, and an inverse branch that activates the inverse-relation neuron. For visual clarity, only the symmetry branch is shown here; the inverse branch follows the same gated-diffusion and binding structure with relation inversion.

Figure 6: Discovery of latent neighborOf on Countries. Blue and gray dashed edges denote the original neighborOf and locatedIn facts; red and orange dashed edges denote the region- and continent-level adjacencies induced by NSR.Region prefixes abbreviate cardinal directions: N (northern), S (southern), E (eastern), W (western), C (central), M (middle), SE (southeastern). Only the $16$ most active countries are shown; all regions and continents are displayed in full.

Explore similar work

Knowledge Graph-Driven Expert-Level Reasoning for Neuroscience
May 24, 2026Jake Stephen, Niraj K. JhaLLM Reasoning Strategies Knowledge Graphs
Department of Electrical and Computer Engineering, Princeton University
Language-encoded network topology enables large language models to reason about complex networks
Sep 3, 2026Ucchwas Talukder Utsha, Sakib Mostafa, James Zou +1Network Science Network Topology
Department of Radiation Oncology, Stanford University, Stanford, California, USA · Department of Electrical Engineering, Stanford University, Stanford, California, USA · Department of Biomedical Data Science, Stanford University School of Medicine, Stanford, California, USA +1
Penelope: Localized Latent Recurrence for Efficient Structured Reasoning
Jul 28, 2026Yutong Chen, Shouqian Shi, Xinran Liu +5Efficient Latent Reasoning Transformer Encoder

May 24, 2026, cs.CLJ/K move · Enter open · S save

Knowledge Graph-Driven Expert-Level Reasoning for Neuroscience

Jake Stephen, Niraj K. Jha

Department of Electrical and Computer Engineering, Princeton University

LLM Reasoning Strategies Knowledge Graphs Reasoning Graphs Neuroscience

Open paper AskPDF arXiv

Knowledge graph (KG) is an abstraction that can be extracted from text corpora and used for in-depth reasoning. Prior work has leveraged KGs to fine-tune language models (LMs), enabling domain-specific superintelligence. In this work, we explore whether KG-driven in-depth reasoning capabilities can emerge in neuroscience using only information contained within a single authoritative textbook. The central hypothesis is that structured knowledge, when distilled into a high-quality KG and converted into KG-grounded question-answer (QA) supervision, is sufficient to produce expert-level reasoning through a fine-tuned LM that surpasses large language models (LLMs) in accuracy, while employing orders of magnitude fewer parameters. We construct a textbook-derived KG via a dual-LLM validation pipeline, expand it with a masked LM trained on the KG topology, generate multi-hop QA items, which include QA pairs and reasoning traces, to fine-tune an LM exclusively on KG-derived supervision, and apply reinforcement learning using path-derived KG signals as implicit reward models. Our results demonstrate that deep, mechanistic neuroscience understanding can be induced in the model without reliance on large, heterogeneous web-scale corpora. The KG-based synthetic neuroscience curriculum that readers can quiz themselves on, and the fine-tuned LM, are available at the following GitHub location: https://kg-bottom-up-superintelligence.github.io/neuro-bench.

Abstract

Explore similar work

Knowledge Graph-Driven Expert-Level Reasoning for Neuroscience

Language-encoded network topology enables large language models to reason about complex networks

Penelope: Localized Latent Recurrence for Efficient Structured Reasoning