NeuroLog: A Family of Neuro-Symbolic Systems

• NeuroLog is a family of research artifacts combining neural, symbolic, and logical techniques to support applications in clinical monitoring, imaging, and software security.
• It integrates smartphone speech biomarkers with relational graph transformers and semantic workflows to provide early detection and precise data fusion.
• Its design emphasizes transparent, auditable intermediate representations and extensibility, merging neural estimation with structured symbolic reasoning.

NeuroLog is not a single standardized research artifact but a recurrent name and naming family attached to several technically distinct systems. In current literature it denotes, among other things, a proposed continuous neurological monitoring platform that combines smartphone speech biomarkers with Relational Graph Transformer architectures, the NeuroLOG/NEUROLOG line of neuro-imaging workflow and neuro-symbolic systems, neural-logic frameworks such as NeuralLog and related lifelong or abstract-machine approaches, and a build-free neuro-symbolic vulnerability-discovery pipeline for C/C++ source code (Norel et al., 4 Dec 2025, Wali et al., 2012, Shah et al., 2022, Guimarães et al., 2021, He et al., 16 Nov 2025, Derkinderen et al., 19 Aug 2025, Rawat, 30 May 2026). This suggests that “NeuroLog” is best understood as a family of research usages rather than a single canonical platform.

1. Major research senses of the name

Across the cited literature, the name appears in several domains with different technical meanings. Some usages are exact project or system names, while others are “NeuroLog-like” architectural blueprints or close naming variants such as NeuralLog, NEUROLOG, and NeuroLOG (Norel et al., 4 Dec 2025, Guimarães et al., 2021, Rawat, 30 May 2026).

Usage	Domain	Core formulation
NeuroLog-like continuous monitoring	Rare neurological disease monitoring	Smartphone speech, PVD, RELGT, risk-stratified alerts
NeuroLOG project	Distributed neuro-imaging infrastructure	OWL-S extensions for jGASW-generated services
NEUROLOG	Hierarchical neuro-symbolic learning	Symbolic predictor, neural extractor, semantic loss
NeuralLog	Neural logic language / NLI	Logic compiled to neural networks; joint neural-logical inference
DeepLog / Neuro-Logic lifelong learning	Neurosymbolic reasoning	Algebraic circuits; shared reusable predicate libraries
NeuroLog	Software security	LLM facts, Soufflé, Z3, likely invariants, ASan

This multiplicity is substantive rather than terminological. In the clinical literature, NeuroLog names a proposal for continuous neurocognitive surveillance; in neuroinformatics, it marks a workflow and service-composition environment; in neural-symbolic AI, it refers to logic-based differentiable systems and their extensions; and in software security, it names an auditable analysis stack that combines LLMs with Datalog and SMT (Norel et al., 4 Dec 2025, Wali et al., 2012, Chen et al., 2021, Rawat, 30 May 2026).

2. Continuous neurological monitoring and digital biomarkers

In the clinical-monitoring sense, NeuroLog is articulated most directly as a system for continuous neurological monitoring that uses everyday smartphone speech as a digital biomarker and fuses it with heterogeneous clinical data through a Relational Graph Transformer (RELGT) (Norel et al., 4 Dec 2025). The motivating claim is that many patients with rare neurological disease report fluctuating cognitive symptoms—especially “brain fog,” working-memory strain, and reduced day-to-day functioning—that remain largely invisible to standard clinic-based neuropsychological testing. The paper frames this as a monitoring gap produced by sparse testing, limited ecological validity, and weak integration with longitudinal symptoms, laboratory data, fatigue, and treatment history.

The most concrete component is the speech pipeline. The proposed capture mode is smartphone-based spontaneous speech, specifically 60-second narratives. From these narratives the system extracts 23 linguistic features and aggregates them into “Proficiency in Verbal Discourse” (PVD), a composite score intended to summarize connected discourse in terms of complexity, detail, coherence, context, and emotion. In the proof-of-concept phenylketonuria cohort, the paper reports $n=42$ individuals with PKU and 41 controls; within PKU, PVD correlated with phenylalanine at $\rho=-0.50, p<0.005$ and tyrosine at $\rho=0.44$, whereas standard cognitive testing showed negligible association, reported as all $r<0.17, p>0.1$ in one passage and all $|r|<0.35$ in the abstract (Norel et al., 4 Dec 2025). The same discussion also notes that approximately 40% were “scoring from speech biomarkers” in a clinically significant way for working-memory deficits, and 45% reported neurocognitive burden, while standard tests remained in the normal range; the paper itself characterizes these percentages as indicative rather than fully established trial-grade evidence.

The RELGT component is presented as the mechanism for contextualizing speech against the relational structure of medical data. The proposed heterogeneous graph includes nodes such as patients, speech sessions, laboratory results, medications, assessments, and symptoms, with edges encoding temporal links, patient ownership, and treatment or causal relations. The envisioned output is a longitudinal prediction engine that issues risk-stratified alerts. The illustrative workflow is explicit: a speech decline at Week 3 triggers additional assessment and reveals metabolic elevation before a scheduled Week 8 blood test, enabling intervention 4+ weeks earlier than standard care; the paper states this as an envisioned workflow rather than a prospectively validated forecasting result (Norel et al., 4 Dec 2025).

Related systems develop complementary modalities within the same broad agenda of digitized neurological assessment. The Digitized Neurological Examination (DNE) platform uses consumer iPhones/iPads to capture finger tapping, finger to finger, forearm roll, and stand-up and walk, extracts 2D or 3D pose, computes clinically interpretable kinematic and spatio-temporal features, and reports classification accuracy beyond 90% for upper-limb tests and 80% for the stand-up and walk tests on a dataset of 21 subjects with normal and simulated-impaired movements (Hoang et al., 2022). NeuroLM treats EEG as a “foreign language,” learns a text-aligned neural tokenizer, pretrains a causal LLM with multi-channel autoregression, scales to 1.7B parameters, and is pretrained on approximately 25,000-hour EEG data for six downstream EEG tasks (Jiang et al., 2024). A distinct conceptual extension appears in Neuro-Linguistic Integration (NLI), where LLMs act as the semantic interface between raw neural data and social application; the paper frames this as a hybrid semantic-neural ecosystem and emphasizes Semantic Transparency, Mental Informed Consent, and Agency Preservation as governing principles (Shenderuk-Zhidkov et al., 18 Mar 2026).

3. NeuroLOG as neuro-imaging workflow and semantic service composition

In the neuro-imaging literature, NeuroLOG designates a distributed resource-sharing environment for brain image datasets and image-processing tools (Wali et al., 2012). Its technical problem is not continuous monitoring but semantic workflow composition in the presence of legacy wrapped applications. The project relied on jGASW, a legacy application wrapper that exposes scientific executables as web services. jGASW-generated WSDL/XSD descriptions are executable but structurally awkward for standard OWL-S tooling, especially because multiple true outputs are hidden inside nested complex types and returned as a single wrapper such as localResult.

The proposed solution is an extension of OWL-S, principally in the Process and Profile layers. The paper introduces the class NlogParameter, defined as a subclass of Parameter and disjoint with Output, to represent parameters embedded inside a composite output. The property nlogExpandsTo links a composite Output to these internal parameters, while hasID stores the identifier needed to locate the corresponding XML tag inside the SOAP result string, and hasLabel stores an informal description (Wali et al., 2012). This is the central modeling repair: the grounding still exposes one monolithic output, but the semantic layer expands that output into typed internal elements that can participate in workflow composition.

The profile layer is linked to the domain ontology OntoNeuroLOG through the property refers-to, allowing a service profile to declare that it performs a domain processing concept such as Registration or De-noising. Workflow dataflow is then expressed with the property links, which connects parameters across services, expanded output components, and workflow-level inputs or outputs. Type compatibility is checked by ontology subsumption. The explicit rule stated in the paper is that a workflow is valid “if Parameters have the same Type or source is subsumed by target according to the dataset ontology,” i.e. if Type(source) = Type(target) or Type(source) \subseteq Type(target) (Wali et al., 2012).

The composition method is semi-automatic. Automatic steps include initial OWL-S generation through WSDL2OWLS, reasoning checks once annotations are present, and automated extraction and transmission of embedded outputs at execution time. Manual or semi-automatic steps include enrichment of the Process model with NlogParameters and IDs, profile annotation with the appropriate OntoNeuroLOG processing class, and user-driven linking of outputs to downstream inputs. Runtime execution uses CORESE queries over the semantic graph to recover nlogExpandsTo, links, hasID, and parameterType, and then parses the returned SOAP envelope accordingly (Wali et al., 2012).

This NeuroLOG line is therefore centered on semantic interoperability, legacy-service composition, and consistency checking. Its significance lies in preserving compatibility with wrapped neuro-imaging software while still supporting semantically typed workflows and ontology-based validation.

4. Neural-logic languages and neurosymbolic reasoning

A separate research line uses the name in the form NeuralLog or in closely related neural-logic frameworks. In “NeuralLog: a Neural Logic Language,” NeuralLog is defined as a first-order logic language that is compiled to a neural network (Guimarães et al., 2021). The program is represented as $(E,F,R)$, with entities, facts, and rules; binary predicates are encoded as matrices, unary predicates as vectors, and propositional predicates as scalars. For a binary predicate matrix $P$, querying with an instantiated first argument represented by one-hot row vector $\mathbf{a}$ yields

$Y = \mathbf{a}P.$

The language also supports numeric attributes and numeric functions, and the paper reports better AUC than comparison systems on Cora and UWCSE for link prediction and on Yelp and PAKDD15 for classification, with comparable results on WordNet (Guimarães et al., 2021).

In natural-language inference, “NeuralLog: Natural Language Inference with Joint Neural and Logical Reasoning” treats NLI as a search problem that starts from the premise and searches for a sequence of valid transformations to the hypothesis (Chen et al., 2021). The system combines a monotonicity-based logical inference engine with neural phrase alignment, uses beam search, and scores states by semantic closeness to the hypothesis. On SICK it reports 90.3% accuracy, and on MED it reports 93.4% accuracy; ablations to -ALBERT-SV and -Monotonicity reduce SICK accuracy to 71.4 and 74.7, respectively (Chen et al., 2021). The technical point is not merely that the system uses logic and neural modules, but that both operate inside a single inference process.

Later neurosymbolic work broadens this idea from task-specific systems to reusable abstract machines and continual predicate libraries. DeepLog is introduced as a neurosymbolic abstract machine with an annotated neural extension of grounded first-order logic and a computational layer based on extended algebraic circuits (Derkinderen et al., 19 Aug 2025). It can express probabilistic, fuzzy, and hybrid semantics, and it compares “logic in the architecture” against “logic in the loss”; on Visual Sudoku 4x4, the probabilistic architecture model reaches $99.71^{+0.04}_{-0.03}\%$ average precision, whereas loss-based logic remains much weaker (Derkinderen et al., 19 Aug 2025). Neuro-Logic Lifelong Learning then recasts neural ILP as a sequence problem with a shared reusable predicate library

$\rho=-0.50, p<0.005$0

implemented on top of Neural Logic Machines; the paper reports forward transfer on arithmetic, tree, graph, and BlocksWorld sequences and shows that replay mitigates catastrophic forgetting (He et al., 16 Nov 2025). In parallel, Neural Logic Networks develop interpretable AND/OR/NOT architectures with biases for unobserved data and factorized IF-THEN rule structures, improving the state of the art in Boolean network discovery and learning compact rules for tabular classification, including a chronic kidney disease example (Perreault et al., 11 Aug 2025).

Taken together, this line of work uses “NeuroLog” in the broader sense of neural-logic learning: logic programs compiled into differentiable networks, hybrid search with logical constraints and neural alignment, and reusable symbolic abstractions learned across tasks.

5. NEUROLOG as abductive neuro-symbolic architecture

In another technically distinct usage, NEUROLOG denotes a hierarchical neuro-symbolic architecture in which a low-level neural extractor produces symbolic features consumed by a high-level symbolic predictor (Shah et al., 2022). The paper formalizes the predictor as a composition $\rho=-0.50, p<0.005$1, where $\rho=-0.50, p<0.005$2 maps raw inputs to symbolic feature assignments and $\rho=-0.50, p<0.005$3 is a symbolic classifier. The key training idea is semantic loss driven by abduction: given an observed class label $\rho=-0.50, p<0.005$4, the symbolic theory enumerates conjunctive feature assignments that would justify that label, and the neural extractor is optimized to assign high probability to at least one such explanation.

The paper’s main contribution is to examine what happens when the original assumption of task-specific feature delineation is removed. In the original chess task, segmentation into 9 board squares already tells the network where each feature is. Without that prior, NEUROLOG can preserve class accuracy while losing explanatory fidelity almost completely. On the chess benchmark, the segmented system $\rho=-0.50, p<0.005$5 reaches 95.41% predictive accuracy and 94.67% explanatory fidelity, whereas the unsegmented from-scratch system $\rho=-0.50, p<0.005$6 reaches 79.88% predictive accuracy but only 0.10% explanatory fidelity. On the synthetic time-series task, $\rho=-0.50, p<0.005$7 reaches 100% / 100%, while $\rho=-0.50, p<0.005$8 reaches 75.79% predictive accuracy and 36.81% explanatory fidelity (Shah et al., 2022). The paper’s succinct conclusion is explicit: “Pre-process. Otherwise pre-train.”

Pre-training is then introduced as feature adaptation: a feature extractor is first trained in a related setting and then adapted with semantic loss alone. This substantially improves fidelity. In chess, $\rho=-0.50, p<0.005$9 raises explanatory fidelity to 62.71%; in time series, pre-trained models reach approximately 75%–76% explanatory fidelity. The authors interpret this as evidence that semantic loss is effective as a constraint-based fine-tuner, but not as a complete substitute for localization priors (Shah et al., 2022).

A later extension, “Neural-Symbolic Integration with Evolvable Policies,” explicitly turns NeuroLog from a framework with a fixed symbolic component into one whose symbolic policy is itself trainable (Thoma et al., 8 Jan 2026). The assumed architecture consists of a NeuralModule, a SymbolicModule, and a predefined Translator. The extension adds induce() to both modules: for the symbolic side, induce() edits the policy by adding or simplifying rules; for the neural side, it performs ordinary gradient-based training. Symbolic policies are represented with Machine Coaching semantics and implemented using Prudens, while neural training still uses abductive supervision with

$\rho=0.44$0

The evolutionary process starts from an empty symbolic policy and random neural weights, applies symbolic mutations $\rho=0.44$1, $\rho=0.44$2, $\rho=0.44$3 and neural mutations $\rho=0.44$4, $\rho=0.44$5, and selects offspring by a relative-fitness scheme. Across 150 experiments, the reported end-of-run performance on test data is Median Correct: 0.992, Mean Correct: 0.944, Median Abstain: 0.000, and Median Wrong: 0.005 (Thoma et al., 8 Jan 2026).

This NEUROLOG lineage is therefore defined by abductive weak supervision, semantic loss, and a persistent concern with the relationship between symbolic correctness and low-level feature extraction.

6. NeuroLog as auditable vulnerability discovery

In software security, NeuroLog names a compile-free vulnerability-analysis pipeline for C/C++ source code that combines LLMs, Datalog, SMT, runtime invariants, and crash synthesis (Rawat, 30 May 2026). Its premise is that analysts currently choose between heavyweight static analyzers that require a working build and free-form LLMs that read source readily but hallucinate and lose track of cross-function dataflow. NeuroLog assigns each layer a sharply delimited role: an LLM extracts typed facts one function at a time; a Soufflé rule mesh composes those facts into cross-function findings; Z3 filters infeasible findings and emits a SAT model for each feasible one; likely range invariants from a few runtime seeds tighten the SMT problem; and a second LLM turns each SAT model into a Python emitter that produces candidate crashing inputs validated by an AddressSanitizer harness.

The fact schema includes relations such as Def, Use, Call, ActualArg, ArithOp, Cast, Guard, FieldRead, FieldWrite, and MemRead, all anchored to concrete source locations. The Datalog layer is organized into five passes—alias, interprocedural taint, type safety, memory safety, and sink—and includes output relations such as TaintedVar, NarrowArithAtSink, TaintedNarrowArith, ImplicitTruncation, SignedArgAtSink, PotentialArithOverflow, UncheckedAlloc, DoubleFree, TypeSafetyFinding, and MemSafetyFinding (Rawat, 30 May 2026). The SMT stage then reconstructs the relevant def-use chain, path guards, and bug condition. For the stb_vorbis allocation-overflow case, the paper gives the bug condition as len + 1 < len over 32-bit unsigned arithmetic, yielding a witness such as len = 0xFFFFFFFF.

The system’s central design claim is auditability. Each stage emits inspectable artifacts: slice selection, typed facts, Datalog derivations, SMT feasible/infeasible decisions, SAT models, generated emitters, and ASan transcripts. The SAT model is preserved as a first-class artifact rather than collapsed into a yes/no verdict, because it also serves as input to crash synthesis (Rawat, 30 May 2026).

Empirically, the paper reports end-to-end re-discovery of eight CVE-class issues across stb, cJSON, libxml2, an FFmpeg demuxer slice, and curl 8.3.0, including the CVSS-9.8 SOCKS5 heap overflow CVE-2023-38545, all ASan-confirmed (Rawat, 30 May 2026). On libarchive HEAD, the pipeline surfaces five memory-safety bugs—four previously unreported—across the cpio reader and the XAR/WARC/7zip writers; all were filed upstream, several fixes were merged, and the cpio use-after-free was acknowledged in seven hours (Rawat, 30 May 2026). The extraction stage takes ~37 s and \$0.005 on stb, the likely-invariant filter from three Matroska seeds eliminates 13.2% of the FFmpeg-demuxer feasible set, and crash synthesis turns a static finding into a 102-byte stb_vorbis crash in two LLM iterations without a fuzzer (Rawat, 30 May 2026).

A related but distinct precursor appears in Log2NS, which also treats observational artifacts as an incomplete view of underlying system behavior and combines learned representations with symbolic reasoning to reduce survivorship bias (Thimmisetty et al., 2021). That earlier framework operates on reactive-system logs rather than source code, but it shares the same general design pattern: neural or statistical structure discovery is used to generate hypotheses, while formal reasoning confirms, refutes, or generalizes them.

7. Cross-cutting themes and unresolved questions

The first recurring theme is that NeuroLog is not one system. Confusion arises because the same or closely related name has been used for a continuous neurocognitive monitor, a neuro-imaging workflow environment, a neural-logic language, an abductive neuro-symbolic classifier, and a vulnerability-discovery stack (Norel et al., 4 Dec 2025, Wali et al., 2012, Guimarães et al., 2021, Shah et al., 2022, Rawat, 30 May 2026). The commonality is therefore architectural rather than disciplinary. This suggests a shared pattern: learned components are repeatedly paired with explicit structured intermediates—graphs, rules, ontologies, Datalog facts, SMT witnesses, or symbolic policies.

A second theme is the preference for inspectable intermediate representations over end-to-end opaque models. The clinical NeuroLog blueprint centers on PVD, patient-centric graphs, and clinician-facing alerts rather than a single raw end score (Norel et al., 4 Dec 2025). NeuroLOG service composition uses NlogParameter, nlogExpandsTo, hasID, and ontology subsumption checks to keep workflow semantics explicit (Wali et al., 2012). NeuralLog and related systems compile predicates and rules into visible neural structures or explicit proof-path searches (Guimarães et al., 2021, Chen et al., 2021). The security NeuroLog stack exposes typed facts, Datalog derivations, SAT models, and ASan traces (Rawat, 30 May 2026).

A third theme is that symbolic correctness and practical deployment remain unresolved. The clinical monitoring proposal is candidly not a complete validation study: it lacks a prospective trial showing that RELGT-based alerts improve outcomes, a direct benchmark against simpler multimodal fusion models, a full specification of the 23 speech features, calibration analysis, multilingual fairness metrics, and live clinical deployment results (Norel et al., 4 Dec 2025). The NeuroLOG service-composition work remains semi-automatic, does not solve automatic discovery and mediation, and knowingly blurs the separation between process and grounding (Wali et al., 2012). In NEUROLOG feature adaptation, semantic loss can preserve end-task prediction while failing badly at recovering the intended low-level symbols; explanatory fidelity is therefore not guaranteed by class accuracy alone (Shah et al., 2022). In lifelong neural-logic learning, catastrophic forgetting appears without replay, and rule interpretability remains limited because Neural Logic Machine predicates are latent rather than fully human-readable (He et al., 16 Nov 2025). In the security setting, LLM recall misses, macro handling, and the compile-free precision gap remain explicit limitations (Rawat, 30 May 2026).

A final theme is extensibility. The clinical NeuroLog proposal is explicitly multi-disease and forward-looking, naming Parkinson’s, Huntington’s, and Wilson’s disease as future targets beyond PKU (Norel et al., 4 Dec 2025). DNE is modular and already proposes expansion to eye movement, facial activation, and phonation (Hoang et al., 2022). Evolvable-policy NeuroLog replaces fixed symbolic rules with mutable ones (Thoma et al., 8 Jan 2026). Security NeuroLog adds optional recon, likely invariants, on-demand Datalog queries, and new rule passes as analysts encounter new bug classes (Rawat, 30 May 2026). The name’s long-term significance therefore lies less in any one implementation than in a persistent research strategy: combining neural estimation with explicit relational, logical, or symbolic structure so that inference remains reviewable, queryable, and, at least in principle, auditable.