Neuro-Symbolic Artificial Intelligence: A Task-Directed Survey in the Black-Box Models Era

AI summary

Overview

Research area: Neuro-Symbolic Artificial Intelligence (NeSy), surveyed through the lens of task-specific applications in the era of large black-box connectionist models, with emphasis on Natural Language Processing (NLP) and Computer Vision (CV).

Technical level: Intermediate. The paper is a survey rather than a new model or benchmark, so it assumes familiarity with neural architectures and with basic symbolic/logical notation, but it surveys work rather than deriving new theory.

One-sentence scope: The paper reviews task-specific neuro-symbolic research to assess how injecting symbolic components into neural pipelines can improve explainability and reasoning, and to characterize when those gains are real versus when they come at a cost.

Note on provenance: the paper states that a definitive, copyrighted, peer-reviewed and edited version is published in IJCAI 2025, pp. 10418-10426, 2025, DOI https://doi.org/10.24963/ijcai.2025/1157. The authors are affiliated with the Department of Computer Science and Engineering, University of Bologna, Cesena Campus (Via dell'Università 50, I-47522 Cesena, Italy). Giovanni Pio Delvecchio and Lorenzo Molfetta are listed as co-first authors with equal contribution.

What This Paper Is About

Since the last AI breakthrough in 2017, connectionist systems have produced unprecedented results, and the field has largely converged on the position that "stacked neural layers is all we need." The paper argues that those advances conceal fundamental drawbacks in data efficiency and explainability. Neuro-symbolic methods promise to infer or exploit behavioral schema and are often treated as one possible proxy for human-level intelligence, but they suffer from limited semantic generalizability and from the difficulty of declining complex domains into pre-defined patterns and rules, which hinders practical deployment in real-world scenarios.

The goal is therefore a task-directed survey: rather than cataloging NeSy methods abstractly, the authors examine task-specific advancements to determine whether and how incorporating symbolic systems actually enhances explainability and reasoning in real-life tasks and applications.

Key Contributions

1. A task-directed framing of the NeSy literature. Instead of organizing the survey around model families or formalisms, the authors organize it around the tasks and applications where neuro-symbolic components are actually used, which is what distinguishes this survey from general NeSy overviews.

2. An assessment of NeSy competitiveness against black-box models. The survey explicitly confronts the question raised by the post-2017 surge in connectionist performance: whether NeSy solutions remain competitive, particularly in NLP and CV.

3. A focus on explainability and reasoning as the motivating benefit. The survey is positioned as a resource for researchers who want explainable NeSy methodologies for real-life tasks, treating explainability and reasoning capability — not raw accuracy alone — as the axis of evaluation.

4. A companion reproducibility repository. Reproducibility details and in-depth comments on each surveyed research work are made available at https://github.com/disi-unibo-nlp/task-oriented-neuro-symbolic.git.

Main Findings

• Black-box dominance masks real weaknesses. The surveyed framing holds that the latest neural models achieve marvelous advancements but conceal fundamental drawbacks regarding data efficiency and explainability, which is what makes trustworthy and efficient solutions a live need.

• NeSy promises schema-level behavior. The ability of NeSy methods to infer or exploit behavioral schema has been widely considered one of the possible proxies for human-level intelligence.

• NeSy has a generalization and domain-encoding bottleneck. Limited semantic generalizability, plus the challenge of declining complex domains with pre-defined patterns and rules, hinders practical implementation in real-world scenarios — this is presented as the central obstacle to adoption.

• Symbolic integration produces mixed, task-dependent results. The survey reports normalized comparisons in which adding symbolic components yields both gains and losses relative to baselines. Reported effects include a +10.8% and a +6.5% improvement in some settings, alongside −20.0%, −10.5%, −10.5% and −3.7% degradations in others, plus a −9.4 change reported on a non-percentage scale. In the available content these deltas are listed alongside baseline figures of 77.9%, 70.3%, 84.4%, 69.0%, 71.5%, 67.0% and 10.6, but the truncated text does not pair each number with the specific task, dataset or method it belongs to, so those pairings cannot be stated here.

• The gain is not automatic. Because the reported effects span strongly positive and strongly negative values, the survey's evidence indicates that symbolic components help on some tasks and hurt on others, which undercuts any blanket claim that NeSy is uniformly superior or uniformly inferior to purely connectionist approaches.

• Coverage spans relational and logical structure. The surveyed material includes relational scoring of the form r(h,t) for knowledge-graph-style reasoning and logical constructs including equivalence and subsumption, indicating that the symbolic side of the surveyed systems ranges from structured relation scoring to explicit logical relations.

Methodology in Plain English

This is a literature survey, so the method is selection and comparison rather than experimentation. The authors gather neuro-symbolic research and organize it by the task it targets — reasoning and language tasks, vision-and-language tasks, and structured/relational tasks — rather than by the formal machinery used. For each surveyed work they record what the neural part does, what the symbolic part contributes, and how the combination performs relative to a neural baseline, using normalized comparison figures to make results across papers comparable. They then layer interpretation on top: where a symbolic injection helped, where it hurt, and what that implies about generalization and domain encoding. To make the survey auditable, they publish reproducibility details and in-depth per-work comments in a public GitHub repository, so a reader can follow the reasoning behind each entry rather than trusting the summary alone.

Important limits on what is reported in the available content: the survey's section-by-section structure, the names of the benchmark datasets and tasks, and the mapping of individual performance figures to individual papers are not present in the content provided, so those specifics are not reported here.

Why This Matters

Impact on research. The survey reframes the NeSy debate. Rather than asking whether neuro-symbolic methods are competitive in the abstract — a question the post-2017 connectionist surge appeared to settle in favor of pure neural models — it asks which tasks actually benefit from symbolic structure and what those benefits cost. The mixed positive and negative deltas it reports make the case that the productive research question is conditional and task-specific, which redirects effort toward identifying the conditions under which symbolic components pay off.

Real-world applications (drawn from the application categories the survey covers):

• NLP systems requiring reasoning over multi-step or structured information, where explainability matters as much as output quality.
• Computer Vision pipelines that must produce interpretable, verifiable decisions rather than opaque classifications.
• Structured and relational data settings, including knowledge-graph-style reasoning over entities and relations, where symbolic relations are a natural fit.
• Trustworthy and data-efficient deployments, where the motivation for NeSy arises precisely from the need for solutions that are both trustworthy and efficient.

Industry relevance. The two drawbacks the survey foregrounds — data efficiency and explainability — are the two that most often block deployment of neural models in regulated or high-stakes settings. Industries that cannot ship an unexplainable decision, or that cannot afford the data volumes a purely connectionist system demands, are the direct audience for the trade-offs this survey quantifies at task level. The negative deltas are as industrially relevant as the positive ones: they warn that bolting symbolic components onto a working neural pipeline is not a free explainability upgrade.

Future Directions

1. Characterize when symbolic injection helps versus hurts. The reported spread from +10.8% and +6.5% down to −20.0%, −10.5% and −9.4 makes predicting the sign of the effect the most pressing open question.

2. Address limited semantic generalizability. The survey identifies this as a core NeSy weakness, pointing to generalization beyond the patterns and rules anticipated at design time as an unsolved problem.

3. Reduce the cost of declining complex domains into rules. The difficulty of encoding pre-defined patterns and rules for complex domains is named as a barrier to practical implementation, so lowering that cost — or removing the need for it — is a natural next step.

4. Extend the task-directed evaluation to more real-life applications. The stated purpose of the survey is to serve researchers building explainable NeSy methodologies for real-life tasks, which implies continued task-level evaluation across domains beyond those surveyed, along with the reproducibility tracking the authors have already begun in their public repository.

Target Audience

Researchers and practitioners working on neuro-symbolic methods, explainable AI, and reasoning systems, particularly those whose focus is NLP or Computer Vision. It is also useful for engineers deciding whether to add symbolic components to an existing neural pipeline, since the survey reports both the gains and the regressions those components can introduce on a per-task basis. Readers looking for new architectures or new experimental results will not find them here; the value is in the task-directed organization, the comparative deltas, and the companion repository of per-work comments, which together serve as a starting map for anyone exploring explainable NeSy approaches for real-life tasks.

Authors’ abstract

The integration of symbolic computing with neural networks has intrigued researchers since the first theorizations of Artificial intelligence (AI). The ability of Neuro-Symbolic (NeSy) methods to infer or exploit behavioral schema has been widely considered as one of the possible proxies for human-level intelligence. However, the limited semantic generalizability and the challenges in declining complex domains with pre-defined patterns and rules hinder their practical implementation in real-world scenarios. The unprecedented results achieved by connectionist systems since the last AI breakthrough in 2017 have raised questions about the competitiveness of NeSy solutions, with particular emphasis on the Natural Language Processing and Computer Vision fields. This survey examines task-specific advancements in the NeSy domain to explore how incorporating symbolic systems can enhance explainability and reasoning capabilities. Our findings are meant to serve as a resource for researchers exploring explainable NeSy methodologies for real-life tasks and applications. Reproducibility details and in-depth comments on each surveyed research work are made available at https://github.com/disi-unibo-nlp/task-oriented-neuro-symbolic.git.