Reasoning Meets Representation: Envisioning Neuro-Symbolic Wireless Foundation Models

Conference: NeurIPS 2025 arXiv: 2511.16369 Code: None Area: Neuro-Symbolic AI / Wireless Communications Keywords: Neuro-Symbolic AI, Wireless Foundation Models, Knowledge Graphs, Differentiable Logic, 6G

TL;DR

This paper proposes a vision framework that integrates the neuro-symbolic (NeSy) paradigm with Wireless Physical-layer Foundation Models (WPFMs)—employing WPFMs as a neural perception engine to generate RF embedding vectors, and an ontology-driven knowledge graph together with a differentiable logic layer as the symbolic reasoning component. The resulting system achieves interpretable, generalizable, and compliance-verifiable wireless AI, providing a concrete technical pathway toward AI-native 6G networks.

Background & Motivation

Background: ML approaches for wireless communications have evolved along two tracks: (1) symbolic methods grounded in domain knowledge (e.g., Maxwell's equations, channel models) that provide transparency and interpretability and can guarantee compliance, but exhibit poor adaptability in high-complexity scenarios and suffer from rule explosion; and (2) neural methods (data-driven deep learning) that learn complex patterns from raw signals such as IQ samples and CSI, yet operate as "black boxes" lacking reliability and trustworthiness. Recently proposed WPFMs achieve general representation transfer across tasks (e.g., from modulation recognition to interference classification, sensing, resource management, and localization) by self-supervised pre-training of Transformers on diverse RF data, representing a significant advance.

Limitations of Prior Work: Even though WPFMs produce powerful general-purpose representations, they inherit the fundamental deficiencies of purely neural networks—lack of interpretability (unable to explain "why this resource block was allocated"), insufficient robustness (fragile under OOD scenarios), limited adaptability (requiring large-scale fine-tuning data), and inability to verify regulatory compliance (e.g., spectrum policies, physical constraints).

Key Challenge: The vision of AI-native 6G networks demands intelligence deeply embedded in the system core with trustworthiness, yet current purely neural paradigms cannot simultaneously satisfy both high performance and high trust.

Goal: To propose a systematic framework that combines WPFMs with symbolic reasoning, endowing wireless AI with both the perceptual capability of neural networks and the reasoning and verification capability of symbolic systems.

Key Insight: The pipeline pattern from the NeSy paradigm is adopted—clearly separating perception from reasoning, with a WPFM generating embeddings on one end and a knowledge-graph- and differentiable-logic-based reasoning engine on the other.

Core Idea: WPFMs generate sub-symbolic embeddings (rather than directly outputting labels); these embeddings are fed into symbolic reasoning modules for interpretable decision-making and constraint verification; end-to-end joint training is enabled through differentiable logic.

Method

Overall Architecture

The framework is a two-module pipeline architecture: (1) the WPFM neural perception engine receives raw RF signals (IQ samples, CIR, CSI, etc.) and generates dense embedding vectors (sub-symbolic representations) via a self-supervised pre-trained Transformer; (2) the symbolic reasoning module receives these embeddings, combines them with domain knowledge stored in ontologies and knowledge graphs, executes inference through a differentiable logic layer, and outputs interpretable decisions. The two modules are connected end-to-end via gradient backpropagation through the differentiable logic layer.

Key Designs

1. WPFM as the Neural Perception Engine:

   • Function: Pre-train a Transformer on diverse RF data to generate reusable embedding representations rather than conventional label classifications.
   • Mechanism: Raw time-series signals (IQ samples, CIR, CSI) serve as input and are tokenized through patching; self-supervised pre-training is performed using masked or next-sample prediction. After pre-training, dense embedding vectors \(\mathbf{z} = f_{\theta}(\mathbf{x}_{\text{RF}})\) are produced as "perceptual input" for the symbolic module. Lightweight fine-tuning suffices to adapt the model to new downstream tasks.
   • Design Motivation: Conventional practice has foundation models directly output classifications or predictions—but this limits the reusability of a single model. Generating embeddings rather than labels allows the output of the same model to be reused by symbolic reasoning modules across different tasks, while providing an intermediate representation interface for interpretable decision-making.

2. Three-Layer Symbolic Reasoning Architecture:

   • Function: Decompose the symbolic reasoning pipeline into three composable components—ontology, knowledge graph, and differentiable logic.
   • Mechanism: (a) The ontology defines the shared semantic schema of the wireless domain—concepts (e.g., "interference source," "resource block") and their relationships; (b) the knowledge graph instantiates the ontology concretely, encoding entities, relations, and attributes for specific scenarios to form a structured world model; (c) the differentiable logic layer replaces discrete logical operations (AND/OR/NOT) with continuous differentiable functions (e.g., t-norms in fuzzy logic), enabling the logical inference process to participate in gradient computation.
   • Design Motivation: The three-layer separation makes the system both flexible and composable—the ontology provides a general semantic framework shared across tasks, the knowledge graph encodes task-specific knowledge (e.g., spectrum policies, physical formulas), and differentiable logic enables end-to-end training so that the symbolic layer is no longer a static rule engine.

3. End-to-End Neuro-Symbolic Joint Training:

   • Function: Enable gradient backpropagation from the symbolic reasoning layer to the WPFM through the differentiable logic layer, supporting joint optimization of the entire hybrid system.
   • Mechanism: Logical rules \(R\) are represented as continuous functions \(\hat{R}\), e.g., \(\text{AND}(a,b) \to T(a,b)\) (t-norm) and \(\text{OR}(a,b) \to S(a,b)\) (t-conorm), transforming the inference process into a differentiable computation graph. The loss function can simultaneously incorporate a task loss (from downstream tasks over WPFM embeddings) and a constraint loss (from compliance verification in the symbolic layer), yielding \(\mathcal{L} = \mathcal{L}_{\text{task}} + \lambda \mathcal{L}_{\text{constraint}}\).
   • Design Motivation: If the symbolic layer is non-differentiable (e.g., traditional Prolog inference), the WPFM and symbolic layer can only be trained separately, precluding the use of reasoning feedback to improve perception. Differentiability unites them into a coherent whole.

Loss & Training

As a vision paper, no concrete loss function implementation or training strategy is provided. The core design principle is that, after logical rules are made continuous via differentiable logic, task loss and constraint verification loss can be optimized end-to-end within the same computation graph using standard gradient descent. Specifically, the differentiable logic layer replaces discrete operations such as AND/OR with continuous t-norms/t-conorms (e.g., product t-norm: \(T(a,b)=a\cdot b\); Łukasiewicz t-norm: \(T(a,b)=\max(0, a+b-1)\)), allowing gradients produced during inference to flow through the entire computation graph. The WPFM self-supervised pre-training phase does not involve the symbolic layer and uses standard masked prediction or next-sample prediction loss. The joint training loss during fine-tuning can be formalized as \(\mathcal{L} = \mathcal{L}_{\text{task}}(\hat{y}, y) + \lambda \mathcal{L}_{\text{constraint}}(R(\mathbf{z}))\), where \(\hat{y}\) is the task prediction and \(R(\mathbf{z})\) is the constraint verification result produced by the symbolic reasoning layer over embedding \(\mathbf{z}\).

Key Experimental Results

Main Results

This is a vision paper and includes no quantitative experimental evaluation. The core contributions are the framework design and the identification of research challenges.

Systematic comparison of three methodological paradigms (Table 1):

Dimension	Symbolic Methods	Neural Methods	Neuro-Symbolic (Ours)
Interpretability	High (explicit rules)	Low (black box)	High (symbolic reasoning)
Scalability	Low (rule explosion)	High (data-driven)	High (inherits neural scalability)
Adaptability	Low (manual modification required)	High (ICL, etc.)	Very High (reasoning + few-shot)
Constraint Compliance	High (hard-coded)	Low (unprovable)	High (symbolically verifiable)
Data Efficiency	High (domain knowledge)	Low (large-scale data required)	Medium (priors reduce data requirements)

Ablation Study

The paper presents three major open research challenges as qualitative analysis:

Challenge	Problem Description	Core Difficulty
Challenge 1	Physical-layer knowledge representation	Unified encoding of continuous physical laws (Maxwell's equations, fading models) and discrete protocol logic (MAC state machines)
Challenge 2	Real-time inference	Under microsecond-level physical-layer constraints, the entire hybrid model must satisfy latency requirements (potentially requiring compilation to FPGA/ASIC)
Challenge 3	Sub-symbolic–symbolic gap	Semantic translation between self-attention's distributed representations and symbolic AI's discrete, localized structures

Key Findings

1. Both symbolic and neural tracks have fatal flaws: symbolic methods cannot scale; neural methods cannot be interpreted—their fusion is the only viable path.
2. "Generating embeddings rather than labels" from WPFMs is a critical design choice: embeddings can be reused by different reasoning modules, whereas labels are task-specific terminal states.
3. Differentiable logic is the core bridging mechanism: without differentiability, end-to-end training is impossible and the symbolic layer degrades into a post-processing filter.
4. Real-time performance is the greatest obstacle to deployment: even if the framework is conceptually feasible, the microsecond-level latency constraints of the physical layer far exceed the inference speed of current NeSy systems.

Highlights & Insights

• Translating "AI-native 6G" from a slogan into a concrete technical pathway—rather than vaguely advocating for "AI," the paper precisely identifies the NeSy paradigm as the required approach, specifies the necessary components (WPFM + ontology + knowledge graph + differentiable logic), and pinpoints three concrete technical bottlenecks.
• The design of "WPFM generating embeddings" carries architectural inspiration—positioning the foundation model as a perception engine rather than a decision-maker leaves room for the upper-layer reasoning module to explain and verify decisions; this idea generalizes to foundation model applications in other domains.
• The identification of the three challenges is precise and well-targeted—the continuous-discrete unification of physical-layer knowledge, hardware compilation for real-time inference, and the sub-symbolic–symbolic gap are core problems that any NeSy system entering the physical world will encounter, not issues unique to the wireless domain.

Limitations & Future Work

• As a purely vision paper without any experimental validation or prototype implementation, the viability of the design choices cannot be assessed—the overall framework remains at the conceptual demonstration stage.
• The construction and maintenance of the knowledge graph are left entirely unaddressed—who encodes the knowledge, how consistency is guaranteed, how the graph auto-updates as network standards evolve (e.g., from 3GPP Release 17 to Release 18), and how semantic heterogeneity across different operators is handled.
• No quantitative analysis of the latency–accuracy Pareto frontier for real-time inference is provided; whether a symbolic reasoning engine can operate with acceptable accuracy under microsecond-level constraints constitutes the greatest technical risk.
• The mechanisms for maintaining cross-layer knowledge consistency (physical layer → MAC → network layer) are ignored; knowledge graphs at different layers may harbor semantic conflicts.
• Failure modes are not discussed: when symbolic reasoning and neural perception yield contradictory results, how arbitration is performed—whether to trust the uncertainty estimates of neural perception or the deterministic conclusions of symbolic reasoning.
• A concrete technical pathway for bridging the sub-symbolic–symbolic gap is absent—how to extract discrete symbols acceptable to the knowledge graph from the distributed attention representations of Transformers remains an open problem.
• Adversarial scenarios are not considered: whether an attacker could manipulate RF inputs to deceive the neural perception engine, thereby misleading symbolic reasoning into producing erroneous yet "interpretable" decisions.

Related Work & Insights

• vs. purely neural WPFM methods (Cheraghinia et al., 2025; Zhou et al., 2025): These works focus on pre-training universal RF representations and achieve breakthroughs in cross-task transfer for modulation recognition and signal classification—but they stop at the perception level and cannot explain decision rationale or verify compliance with spectrum policies. This paper adds reasoning and verification capabilities on top of that foundation.
• vs. traditional symbolic methods (Bhuyan et al., 2024): Traditional symbolic methods are constrained in complexity and adaptability—the number of rules grows exponentially with scenario complexity (rule explosion), and rules must be manually edited for each new scenario. The neural perception engine in this paper supplements the ability to automatically learn complex patterns from data.
• vs. general NeSy frameworks (Schwalbe et al., 2024): General NeSy research offers paradigmatic guidance but does not address the specific constraints of wireless physical layers (microsecond-level latency, spectrum compliance, encoding of continuous physical laws, real-time hardware deployment). This paper is the first to instantiate the NeSy framework specifically for the wireless domain and identifies domain-specific technical bottlenecks.
• vs. AI-native 6G frameworks (10273257): The 6G vision literature extensively discusses the concept of "AI-native" networks but lacks concrete AI architecture design. This paper fills the gap from concept to architecture by adopting WPFM + NeSy as a specific technical choice.
• Insights: The application of differentiable logic in the wireless domain may inspire broader "physics-constrained AI" design—any domain requiring simultaneous data-driven performance and physical/regulatory compliance (e.g., autonomous driving, medical AI, industrial control) can draw on the NeSy paradigm.

Rating

• Novelty: ⭐⭐⭐⭐ — Systematically applying the NeSy paradigm to wireless foundation models is a first; the identification of the three challenges is insightful.
• Experimental Thoroughness: ⭐⭐ — Purely a vision paper with no experiments or prototype validation.
• Writing Quality: ⭐⭐⭐⭐ — The argumentation is logically clear, the comparison in Table 1 is effective, and the research challenges are described precisely.
• Value: ⭐⭐⭐ — High directional value but distant from practical deployment; inspiring for both the wireless AI and NeSy communities, yet lacking actionable guidance.

Related Papers

• [NeurIPS 2025] Deliberation on Priors: Trustworthy Reasoning of Large Language Models on Knowledge Graphs
• [NeurIPS 2025] GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation
• [CVPR 2026] Mario: Multimodal Graph Reasoning with Large Language Models
• [NeurIPS 2025] From Sequence to Structure: Uncovering Substructure Reasoning in Transformers
• [NeurIPS 2025] The Underappreciated Power of Vision Models for Graph Structural Understanding