# Neuro-symbolic AI

> Source: https://aiwiki.ai/wiki/neuro_symbolic_ai
> Updated: 2026-11-18
> Fact-checked: 2026-11-18
> Categories: AI Research, Artificial Intelligence, Machine Learning, Neural Networks
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Neuro-symbolic AI." aiwiki.ai, 18 Nov 2026. https://aiwiki.ai/wiki/neuro_symbolic_ai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Neuro-symbolic AI** (also written neurosymbolic AI, neural-symbolic AI or NeSy AI) is the branch of [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) that combines [neural networks](https://aiwiki.ai/wiki/neural_network), which learn from data, with symbolic methods such as formal logic, rules and search, which represent knowledge explicitly and reason over it.[1][2] Its proponents argue that the pairing can join the strengths of the two traditions: trainability from raw data and tolerance of noise on the neural side, and explainability, provable correctness and the use of expert knowledge on the symbolic side.[2] Well-known examples range from early systems that compiled rules into network weights, through logic-based training objectives such as Logic Tensor Networks and probabilistic logic programs with neural components such as DeepProbLog, to [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind)'s [AlphaGeometry](https://aiwiki.ai/wiki/alphageometry), which pairs a neural language model with a symbolic deduction engine to prove olympiad geometry theorems.[3][10][12][20]

The field is older than [deep learning](https://aiwiki.ai/wiki/deep_learning). A dedicated workshop series has met since 2005, and research output at the major machine learning conferences rose sharply after about 2016.[2][6] Interest grew further around 2020, when researchers including [Gary Marcus](https://aiwiki.ai/wiki/gary_marcus) argued that deep learning alone would not yield robust AI and called for hybrid systems.[1][7] As of November 2026, the label covers a wide range of systems, from differentiable logics to [large language models](https://aiwiki.ai/wiki/large_language_model) that hand work to symbolic solvers.[26][27] Critics such as [Richard Sutton](https://aiwiki.ai/wiki/richard_sutton) have argued that general methods that scale with computation tend to outperform approaches that build in human knowledge.[30]

## Background

[Symbolic AI](https://aiwiki.ai/wiki/symbolic_ai) and [connectionism](https://aiwiki.ai/wiki/connectionism) differ first in how they represent information. A symbolic system stores knowledge explicitly, in a form a person can read and manipulate, such as the rule square(x) → rectangle(x). A neural system stores it in weighted connections and patterns of activation spread over many units, sometimes called subsymbolic or distributed representations, which a human observer cannot readily interpret.[2] Artur d'Avila Garcez and Luis Lamb frame this as a contrast between localist representations, in which each concept has its own identifier, and distributed ones, in which concepts are vectors of continuous values; they argue that neuro-symbolic systems need a bridge between the two.[1]

Surveys of the field list similar hoped-for benefits. Kansas State University researchers Md Kamruzzaman Sarker, Lu Zhou, Aaron Eberhart and Pascal Hitzler wrote in 2021 that combining symbolic methods with deep learning might improve out-of-vocabulary handling, training from small data sets, recovery from errors and explainability, compared with systems that rely on deep learning alone.[2] Garcez and colleagues add that symbolic knowledge can act as a constraint on learning that eases problems such as catastrophic forgetting and poor extrapolation, while neural learning can reduce the brittleness of purely symbolic systems.[3]

The word "symbolic" is used in a broad and a narrow sense. Broadly, it covers any approach based on explicit symbol manipulation, including term rewriting, graph algorithms and question answering; narrowly, it means methods based on formal logic, as used in [knowledge representation](https://aiwiki.ai/wiki/knowledge_representation) and reasoning.[2] Spelling also varies. Garcez and Lamb note that the research community long used "neural-symbolic", that "neuro-symbolic" became common in publications and the press, and they adopted the single word "neurosymbolic" to mark what they described as the coming of age of a new research area.[1]

<figure class="embedvideo">
<iframe src="https://www.youtube-nocookie.com/embed/ZfWDVO3rzeA" title="What Is NeuroSymbolic AI? Bridging Reasoning &amp; Neural Networks"></iframe>
<figcaption>"What Is NeuroSymbolic AI? Bridging Reasoning &amp; Neural Networks", video from IBM Technology, published February 23, 2026. <a href="https://www.youtube.com/watch?v=ZfWDVO3rzeA">Watch on YouTube</a>.</figcaption>
</figure>

## History

### Early roots

Sarker and colleagues trace the field's roots to Warren McCulloch and Walter Pitts's 1943 paper on a logical calculus of nervous activity, which predates AI as a discipline. They note that much relevant work was never labeled neuro-symbolic and that the term came into more systematic use only from the 1990s.[2] A long-standing limitation of neural networks in this setting was that they could represent propositional logic but struggled with richer logics; John McCarthy called this the "propositional fixation" of neural networks as early as 1988.[2]

Garcez and colleagues name two systems from the 1990s among the most influential early models that combine logical reasoning and neural learning. The Knowledge-Based Artificial Neural Network (KBANN), published by Geoffrey Towell and Jude Shavlik in 1994, represented if-then rules as stacked perceptrons whose weights were set so that ordinary feedforward computation reproduced the rules' inferences. Garcez and colleagues describe it as the first system to allow learning with background knowledge in neural networks together with knowledge extraction, with applications in bioinformatics.[3][4] The Connectionist Inductive Learning and Logic Programming (CILP) system of Garcez and Gerson Zaverucha, published in 1999, drew inspiration from KBANN. It showed that networks with semi-linear neurons can approximate the fixed-point operator of propositional logic programs with negation, which allowed logic programs to serve as background knowledge for networks trained by [backpropagation](https://aiwiki.ai/wiki/backpropagation).[3][5] In the early 2000s, ensembles of CILP networks were extended to modal and temporal logics, and were used to learn solutions to multi-agent reasoning problems such as the muddy children puzzle.[3]

### A research community

The International Workshop on Neural-Symbolic Learning and Reasoning (NeSy) first met in 2005. At its tenth edition, held as Dagstuhl seminar 14381 in Wadern, Germany, in September 2014, participants decided to turn the series into an association with a charter and formal governance. The association now runs an annual conference on Neurosymbolic Learning and Reasoning, and its website lists the journal *Neurosymbolic Artificial Intelligence* among the community's outlets.[6] That journal is published by IOS Press, part of SAGE Publishing, under a gold open-access model with articles licensed CC BY 4.0.[32]

Sarker and colleagues counted original neuro-symbolic research papers at five leading venues (ICML, NeurIPS, AAAI, IJCAI and ICLR) from 2011 to 2020. They found 43 in total, only one of them before 2016 and 40 from 2017 onward, which they read as a shift toward mainstream adoption, while noting that the totals were still small compared with the number of papers those conferences accept.[2]

### The 2020 debate

The deep learning boom brought both success and criticism. Garcez and Lamb summarize the criticism as brittleness (susceptibility to adversarial attacks), a lack of explainability and a lack of parsimony, meaning heavy demands for data, computation and energy.[1] They describe the AAAI-20 conference, held in New York in February 2020, as a focal point of this debate.[1][35] In a fireside chat with [Geoffrey Hinton](https://aiwiki.ai/wiki/geoffrey_hinton), [Yoshua Bengio](https://aiwiki.ai/wiki/yoshua_bengio) and [Yann LeCun](https://aiwiki.ai/wiki/yann_lecun), psychologist Daniel Kahneman said he thought his intuitive "System 1" understands language while the deliberative "System 2" involves some manipulation of symbols, a distinction many researchers have mapped onto deep learning and symbolic reasoning respectively.[1] And computer scientist Henry Kautz used his Robert S. Engelmore Memorial Lecture, "The Third AI Summer", delivered on 10 February 2020, to propose a taxonomy of ways to combine neural and symbolic methods (see below).[1][2][35]

The same month, Marcus posted "The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence", which proposed "a hybrid, knowledge-driven, reasoning-based approach, centered around cognitive models" in place of the field's emphasis on ever-larger training sets and more compute.[7] In a March 2022 essay for *Nautilus*, "Deep Learning Is Hitting a Wall", he argued that hybrid AI, rather than deep learning or symbols alone, was the best way forward, and pointed to [AlphaGo](https://aiwiki.ai/wiki/alphago)'s use of tree search and to [AlphaFold](https://aiwiki.ai/wiki/alphafold) 2 as systems that already mixed the two.[8] In their 2020 paper "Neurosymbolic AI: The 3rd Wave", later published in *Artificial Intelligence Review*, Garcez and Lamb presented neuro-symbolic integration as a basis for the next wave of AI systems and linked it to questions of trust, safety, interpretability and accountability.[1]

## Kautz's taxonomy

Kautz's taxonomy is a common way of sorting neuro-symbolic systems, and both surveys cited here use it. Kautz's own written version of the lecture, published in *AI Magazine* in 2022, describes six designs and gives each a name, and Garcez and Lamb also describe six types.[1][35] Sarker and colleagues, working from the recorded lecture, list five categories, with partly different notation: "Neuro → Symbolic" for the third design and "Neuro ∪ compile(Symbolic)" for the fourth.[2] The table follows Kautz's numbering and names.

| Type | Kautz's name | Description | Examples cited |
|---|---|---|---|
| 1 | symbolic Neuro symbolic | Standard deep learning whose inputs and outputs are symbols, such as text | Language translation, question answering |
| 2 | Symbolic[Neuro] | A symbolic problem solver that calls a neural network as a subroutine | AlphaGo with [Monte Carlo tree search](https://aiwiki.ai/wiki/monte_carlo_tree_search), [AlphaZero](https://aiwiki.ai/wiki/alphazero) |
| 3 | Neuro \| Symbolic | A neural network for one task (for example object detection) passes its output to a symbolic system for a complementary task (for example query answering) | Neuro-Symbolic Concept Learner, DeepProbLog |
| 4 | Neuro: Symbolic → Neuro | Symbolic knowledge is compiled into the training data or the network itself | Deep learning for symbolic mathematics; Garcez and Lamb also place Logical Neural Networks here |
| 5 | Neuro<sub>Symbolic</sub> | Tightly coupled, distributed systems in which logical rules become embeddings that act as soft constraints on the loss | Logic Tensor Networks, tensor product representations |
| 6 | Neuro[Symbolic] | True symbolic reasoning inside a neural engine, including combinatorial reasoning, possibly using an attention schema | None fully fledged for combinatorial reasoning, as of 2020 |

Sources: Kautz (names), Garcez and Lamb (descriptions, examples) and Sarker et al. (examples).[1][2][35] In Kautz's example for type 4, Guillaume Lample and François Charton trained sequence-to-sequence models on generated datasets for symbolic integration and differential equations, and reported results that beat the commercial computer algebra systems Matlab and Mathematica.[31]

Sarker and colleagues applied the categories to their 43 papers and found the first four reasonably balanced, with no papers in the final category, which they note Kautz presented as a forward-looking goal. They also found the categories ambiguous in practice: assigning papers to them "was not straightforward at all".[2] Garcez and Lamb argue that the tension between effective learning and sound reasoning may favor hybrid designs in the middle of the scale (types 3 to 5).[1]

## Approaches and representative systems

The systems below are grouped by how they connect the neural and symbolic parts. Many fit more than one group.

| System | Year | Integration | Developers |
|---|---|---|---|
| KBANN | 1994 | Rules compiled into network weights | Towell, Shavlik |
| CILP | 1999 | Logic programs compiled into recurrent networks | Garcez, Zaverucha |
| Neural Module Networks | 2015 | Question structure assembles neural modules | Andreas et al. |
| Logic Tensor Networks | 2016 | First-order fuzzy logic as a training objective | Serafini, Garcez |
| Neural Theorem Provers | 2017 | Differentiable backward chaining | Rocktäschel, Riedel |
| Differentiable ILP | 2017 | Rule learning by gradient descent | Evans, Grefenstette |
| Semantic loss | 2017 | Propositional constraint as a loss term | Xu, Van den Broeck et al. |
| DeepProbLog | 2018 | Neural predicates in a probabilistic logic program | Manhaeve, De Raedt et al. |
| NS-VQA | 2018 | Scene parsing plus symbolic program execution | Yi, Wu et al. |
| NS-CL | 2019 | Learned concepts executed by a symbolic program | Mao et al. |
| Logical Neural Networks | 2020 | Neurons correspond to logical formula components | Riegel, Gray et al. |
| PAL | 2022 | Language model writes code, interpreter runs it | Gao, Madaan, Zhou et al. |
| Scallop | 2023 | Differentiable Datalog-based language | Li, Huang, Naik |
| Logic-LM | 2023 | Language model formalizes, symbolic solver infers | Pan et al. |
| AlphaGeometry | 2024 | Language model guides a symbolic deduction engine | Google DeepMind (Trinh et al.) |

Years are those of the first public paper or preprint.[3][4][5][9][10][12][13][14][15][16][17][18][19][20][25][26]

### Compiling knowledge into networks

The oldest strategy translates symbolic knowledge into a network's architecture and initial weights. KBANN and CILP work this way: the rules fix which neurons connect and how strongly before training begins, and learning then refines the encoded knowledge from examples.[1][3] Garcez and Lamb describe this as a "neural-symbolic cycle" in which logic acts as a compiler for the network and rule extraction as a decompiler that reads symbolic descriptions back out of a trained model.[1] Logical Neural Networks, proposed by Ryan Riegel, Alexander Gray and colleagues in 2020, push the correspondence further: every neuron stands for a component of a formula in a weighted real-valued logic, inference can run in any direction rather than toward fixed outputs, and training minimizes a loss that measures logical contradiction.[14]

### Logic as a loss function

A second strategy leaves the network architecture alone and turns logical knowledge into a differentiable penalty. The semantic loss of Jingyi Xu, Guy Van den Broeck and colleagues (ICML 2018) applies to a propositional sentence $\alpha$ over variables $X_1, \dots, X_n$ whose probabilities $p_i$ are the outputs of a network:[9]

$$L^s(\alpha, p) \propto -\log \sum_{x \models \alpha} \; \prod_{i:\, x \models X_i} p_i \prod_{i:\, x \models \neg X_i} (1 - p_i)$$

The sum runs over all assignments $x$ that satisfy the constraint, so the loss is the negative log probability that sampling each output independently according to $p$ produces a state satisfying $\alpha$; it is zero when the outputs already satisfy the constraint with certainty. The authors reported near state-of-the-art results in semi-supervised classification and better prediction of structured objects such as rankings and paths.[9]

Logic Tensor Networks (LTN), introduced by Luciano Serafini and Garcez in 2016, extend the idea to first-order logic. Their language, Real Logic, gives formulas truth values in the interval $[0, 1]$ and interprets constants as vectors of real numbers; the 2016 paper implemented it as deep tensor neural networks using TensorFlow primitives.[10] A 2022 paper in *Artificial Intelligence* by Samy Badreddine, Garcez, Serafini and Michael Spranger notes that many fuzzy-logic operators cause vanishing or exploding gradients and favors a "product" configuration. For truth values $a, b \in [0,1]$, negation is $1 - a$, conjunction is the product $ab$, disjunction is $a + b - ab$, and implication is $1 - a + ab$. The universal quantifier is approximated by a generalized mean of the errors, a smooth minimum,

$$A_{pME}(a_1, \dots, a_n) = 1 - \left( \frac{1}{n} \sum_{i=1}^{n} (1 - a_i)^p \right)^{1/p}, \quad p \ge 1,$$

and the existential quantifier by an ordinary generalized mean, a smooth maximum. Learning searches for the network parameters that maximize the aggregated satisfaction of all the formulas in a knowledge base.[11] The paper shows the same framework handling clustering, multi-label classification, relational learning, query answering, semi-supervised learning, regression and embedding learning, with examples written in [TensorFlow](https://aiwiki.ai/wiki/tensorflow) 2.[11] The reference implementation is published on GitHub under the MIT license, and a separate [PyTorch](https://aiwiki.ai/wiki/pytorch) implementation, LTNtorch, is also available.[33][34] Garcez and Lamb note one consequence of this design: because the logic sits in the loss rather than in the network, the trained network and the logic act as communicating modules, and theorem proving is left to the symbolic side.[1]

### Neural predicates in logic programs

The reverse arrangement keeps a logic program in charge and lets neural networks supply some of its facts. DeepProbLog, introduced by Robin Manhaeve, Luc De Raedt and colleagues (NeurIPS 2018), extends the probabilistic logic programming language ProbLog with "neural predicates" whose probabilities come from neural networks.[12] One demonstration in the paper is MNIST addition: given only pairs of handwritten digit images and the sum of the two digits, the program states the rule that the sum of two images is the sum of the digits they show, and the only learned component is the network that classifies each [MNIST](https://aiwiki.ai/wiki/mnist) image. The digit classifier is never trained directly; gradients from the sum labels flow back through the logic, and the authors report that the trained program extends to multi-digit numbers without further training.[12]

Scallop, from Ziyang Li, Jiani Huang and Mayur Naik (PLDI 2023), packages this style as a programming language. It combines a relational data model, a declarative logic language based on Datalog with recursion, aggregation and negation, and differentiable reasoning built on provenance semirings. The authors evaluated it on eight neuro-symbolic applications and reported accuracy comparable to or better than the models they compared against, with gains in runtime, data efficiency, interpretability and generalization.[13]

### Differentiable reasoning and rule learning

Some systems make the reasoning step itself differentiable. Neural Theorem Provers, from Tim Rocktäschel and Sebastian Riedel (NIPS 2017), recursively build a network modeled on [Prolog](https://aiwiki.ai/wiki/prolog)'s backward-chaining algorithm and replace exact symbolic unification with a similarity kernel over learned vector representations of symbols. Trained by gradient descent on incomplete knowledge bases, the model learned to induce function-free first-order rules and outperformed the link-prediction model ComplEx on three of four benchmark knowledge bases.[15]

Richard Evans and Edward Grefenstette's Differentiable Inductive Logic Programming (∂ILP), published in the *Journal of Artificial Intelligence Research* in 2018, takes a similar route for rule learning. Traditional inductive logic programming learns rules from very little data but breaks on noisy or mislabeled inputs and cannot work on raw data such as pixels; ∂ILP is trained by backpropagation, tolerates noise, and can be connected to neural networks that process ambiguous inputs.[16]

### Neural perception with symbolic execution

In visual question answering, several systems split perception from reasoning. Neural Module Networks (Jacob Andreas and colleagues, 2015) parse a question into its linguistic structure and use it to assemble a network from reusable, jointly trained neural modules for tasks such as recognizing objects or classifying colors.[17] The Neural-Symbolic VQA system (NS-VQA) of Kexin Yi, Jiajun Wu and colleagues (NeurIPS 2018) goes further: it recovers a structured representation of the scene from the image and a program from the question, then executes the program symbolically on the scene representation. It reached 99.8% accuracy on the CLEVR benchmark, and the authors highlighted its data efficiency and the transparency of each reasoning step.[18] The Neuro-Symbolic Concept Learner (Jiayuan Mao and colleagues, ICLR 2019) removed the need for explicit supervision of concepts or programs, learning visual concepts, words and semantic parsing from images paired with questions and answers.[19]

### Neural guidance for symbolic search

In Kautz's type 2, a symbolic search procedure does the reasoning and a neural network guides it. Kautz cited AlphaGo and AlphaZero, which combine neural networks with tree search.[1][2] Several recent examples come from mathematics. DeepMind's AlphaGeometry, published in *Nature* in January 2024 by Trieu Trinh, Yuhuai Wu, Quoc Le, He He and Thang Luong, is described by its authors as a neuro-symbolic system: a neural language model, trained from scratch on synthetic data, guides a symbolic deduction engine through the many branching points of a geometry proof.[20] DeepMind said the final training set held 100 million unique synthetic examples and compared the design to the idea of "thinking, fast and slow": the language model quickly suggests auxiliary constructions such as new points, lines or circles, and the symbolic engine applies formal rules to draw conclusions in a slower, explainable way.[21] On 30 recent olympiad geometry problems, AlphaGeometry solved 25 within competition time limits. The previous best method, Wu's method, solved 10, and the average human gold medalist solved 25.9.[20][21] DeepMind released the code and model publicly.[21]

[AlphaGeometry 2](https://aiwiki.ai/wiki/alphageometry_2) replaced the language model with one based on the [Gemini](https://aiwiki.ai/wiki/gemini) architecture, extended the formal geometry language, and sped up the symbolic engine. Its authors report that it surpassed an average gold medalist at solving olympiad geometry problems, and that it raised the share of [International Mathematical Olympiad](https://aiwiki.ai/wiki/international_mathematical_olympiad) geometry problems from 2000 to 2024 that the language could express from 66% to 88%, and the share solved from 54% to 84%.[23] At the 2024 IMO it solved Problem 4 within 19 seconds of receiving the formalized statement, and together with [AlphaProof](https://aiwiki.ai/wiki/alphaproof) it solved four of six problems for 28 of 42 points, the standard of a silver medal.[22] AlphaProof itself is an AlphaZero-inspired reinforcement learning agent that learns to find formal proofs in the [Lean](https://aiwiki.ai/wiki/lean) language; its *Nature* paper appeared in November 2025.[22][24] DeepMind described AlphaGeometry 2 as a "neuro-symbolic hybrid system".[22] A year later the picture shifted. At IMO 2025, DeepMind reported that an advanced version of Gemini Deep Think solved five of the six problems for 35 points, a gold-medal score, working end-to-end in natural language within the 4.5-hour time limit, whereas the 2024 systems had needed experts to translate problems into formal languages such as Lean and took two to three days. DeepMind said it was continuing work on its formal systems and expected agents that combine natural language with verified formal reasoning to become useful tools.[36]

### Language models with symbolic tools

Since about 2022, several systems have paired [large language models](https://aiwiki.ai/wiki/large_language_model) with symbolic executors. Program-aided Language Models (PAL), from Luyu Gao, Aman Madaan, Shuyan Zhou and colleagues, prompt a model to translate a word problem into a program and leave the computation to a Python interpreter. Using [OpenAI Codex](https://aiwiki.ai/wiki/openai_codex), PAL beat [PaLM](https://aiwiki.ai/wiki/palm)-540B with [chain-of-thought](https://aiwiki.ai/wiki/chain_of_thought) prompting on the [GSM8K](https://aiwiki.ai/wiki/gsm8k) math benchmark by 15 percentage points of top-1 accuracy.[25] Logic-LM (Liangming Pan and colleagues, Findings of EMNLP 2023) has a language model translate a natural-language problem into a symbolic formulation for a deterministic solver, and uses the solver's error messages to repair faulty formalizations. Across five logical reasoning datasets it reported average gains of 39.2% over standard prompting and 18.4% over chain-of-thought prompting.[26]

Some researchers now treat such [tool use](https://aiwiki.ai/wiki/tool_use) as part of the field. An August 2026 preprint by 17 authors, including Garcez, Lamb and Frank van Harmelen, proposes four design principles (Reasoning, Assurances, Interfacing and Learning, abbreviated RAIL) and argues that many leading systems not usually called neuro-symbolic, from physics-aware machine learning and the neural-guided search of DeepMind's "Alpha" systems to tool-augmented language models, can be analyzed in these terms.[27]

## Government and industry programs

The US [Defense Advanced Research Projects Agency](https://aiwiki.ai/wiki/darpa) runs a program called Assured Neuro Symbolic Learning and Reasoning (ANSR). It seeks hybrid algorithms that integrate symbolic reasoning with data-driven learning, together with an assurance framework that gathers different kinds of evidence for safety and risk assessments. The program defines a trustworthy system in part as one that is robust to domain-informed and adversarial perturbations. Its program manager is Susmit Jha.[28]

In industry, [Amazon](https://aiwiki.ai/wiki/amazon) Distinguished Scientist and vice president Byron Cook has argued that trust in AI systems requires neurosymbolic AI, which he described, in an interview with Amazon CTO Werner Vogels republished on Vogels's blog in February 2026, as, generally speaking, the combination of symbolic and statistical methods. In the same interview he pointed to the automated reasoning checks feature of Amazon Bedrock Guardrails, which lets customers try to prove that their own AI outputs are correct, and claimed it could verify accuracy "by up to 99%".[29]

## Debate and criticism

The case for neuro-symbolic AI is contested. A frequently cited counterargument is Sutton's 2019 essay "The Bitter Lesson", which concludes from 70 years of AI research that "general methods that leverage computation are ultimately the most effective, and by a large margin." Sutton argues that building human knowledge into systems helps in the short term but eventually plateaus and can inhibit progress, while breakthroughs come from search and learning methods that scale with computation; in his words, "building in how we think we think does not work in the long run."[30] Supporters of hybrid systems answer that some of the most successful systems already combine both: Marcus cites AlphaGo's tree search, an idea from the late 1950s, as a symbolic component, and in 2022 he pointed to a NetHack competition in which a purely symbolic system outscored the best deep learning entries.[8]

Within the field, researchers acknowledge open problems. Garcez and Lamb describe a tension between principled integration and practical value. They note that unification, the algorithm that computes symbolic substitutions in logical reasoning, is notoriously hard to implement inside neural networks, and that the most practical option may be to compute it symbolically.[1] They also argue that neural networks can represent propositional logic and fragments of first-order logic but not full first-order or higher-order logic, which is part of why Logic Tensor Networks place the logic in the loss function.[1] Among the papers Sarker and colleagues surveyed, few addressed out-of-distribution data or learning from small data, and hardly any addressed error recovery, although these are among the field's claimed benefits.[2] The scope of the label is also debated. The RAIL authors argue that combining machine learning with formal reasoning is not a niche approach but already underlies many successful techniques, including some not usually described as neuro-symbolic.[27]

## See also

- [Symbolic AI](https://aiwiki.ai/wiki/symbolic_ai)
- [Connectionism](https://aiwiki.ai/wiki/connectionism)
- [Expert system](https://aiwiki.ai/wiki/expert_system)
- [Knowledge graph](https://aiwiki.ai/wiki/knowledge_graph)
- [Automated theorem proving](https://aiwiki.ai/wiki/automated_theorem_proving)
- [Explainable AI](https://aiwiki.ai/wiki/explainable_ai)
- [AlphaGeometry](https://aiwiki.ai/wiki/alphageometry)

## References

1. Artur d'Avila Garcez, Luis C. Lamb. "Neurosymbolic AI: The 3rd Wave". arXiv:2012.05876, 10 December 2020; published in Artificial Intelligence Review 56(11):12387-12406, 2023. https://arxiv.org/abs/2012.05876
2. Md Kamruzzaman Sarker, Lu Zhou, Aaron Eberhart, Pascal Hitzler. "Neuro-Symbolic Artificial Intelligence: Current Trends". arXiv:2105.05330, May 2021. https://arxiv.org/abs/2105.05330
3. Artur d'Avila Garcez, Marco Gori, Luis C. Lamb, Luciano Serafini, Michael Spranger, Son N. Tran. "Neural-Symbolic Computing: An Effective Methodology for Principled Integration of Machine Learning and Reasoning". arXiv:1905.06088, 15 May 2019. https://arxiv.org/abs/1905.06088
4. Geoffrey G. Towell, Jude W. Shavlik. "Knowledge-based artificial neural networks". Artificial Intelligence 70(1-2):119-165, October 1994. https://doi.org/10.1016/0004-3702(94)90105-8
5. Artur S. Avila Garcez, Gerson Zaverucha. "The Connectionist Inductive Learning and Logic Programming System". Applied Intelligence 11(1):59-77, July 1999. https://doi.org/10.1023/A:1008328630915
6. Neurosymbolic AI (NeSy) Association. "About the NeSy Association". nesy-ai.org, accessed 18 November 2026. https://nesy-ai.org/
7. Gary Marcus. "The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence". arXiv:2002.06177, 14 February 2020. https://arxiv.org/abs/2002.06177
8. Gary Marcus. "Deep Learning Is Hitting a Wall". Nautilus, 10 March 2022. https://nautil.us/deep-learning-is-hitting-a-wall-238440/
9. Jingyi Xu, Zilu Zhang, Tal Friedman, Yitao Liang, Guy Van den Broeck. "A Semantic Loss Function for Deep Learning with Symbolic Knowledge". Proceedings of the 35th International Conference on Machine Learning (ICML 2018); arXiv:1711.11157. https://arxiv.org/abs/1711.11157
10. Luciano Serafini, Artur d'Avila Garcez. "Logic Tensor Networks: Deep Learning and Logical Reasoning from Data and Knowledge". arXiv:1606.04422, 14 June 2016. https://arxiv.org/abs/1606.04422
11. Samy Badreddine, Artur d'Avila Garcez, Luciano Serafini, Michael Spranger. "Logic Tensor Networks". Artificial Intelligence 303:103649, February 2022; arXiv:2012.13635. https://arxiv.org/abs/2012.13635
12. Robin Manhaeve, Sebastijan Dumančić, Angelika Kimmig, Thomas Demeester, Luc De Raedt. "DeepProbLog: Neural Probabilistic Logic Programming". NeurIPS 2018; arXiv:1805.10872. https://arxiv.org/abs/1805.10872
13. Ziyang Li, Jiani Huang, Mayur Naik. "Scallop: A Language for Neurosymbolic Programming". Proceedings of the ACM on Programming Languages 7(PLDI):1463-1487, June 2023. https://doi.org/10.1145/3591280
14. Ryan Riegel, Alexander Gray, Francois Luus, et al. "Logical Neural Networks". arXiv:2006.13155, 23 June 2020. https://arxiv.org/abs/2006.13155
15. Tim Rocktäschel, Sebastian Riedel. "End-to-End Differentiable Proving". NIPS 2017; arXiv:1705.11040. https://arxiv.org/abs/1705.11040
16. Richard Evans, Edward Grefenstette. "Learning Explanatory Rules from Noisy Data". Journal of Artificial Intelligence Research 61:1-64, 2018. https://doi.org/10.1613/jair.5714
17. Jacob Andreas, Marcus Rohrbach, Trevor Darrell, Dan Klein. "Neural Module Networks". arXiv:1511.02799, November 2015. https://arxiv.org/abs/1511.02799
18. Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, Joshua B. Tenenbaum. "Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding". NeurIPS 2018; arXiv:1810.02338. https://arxiv.org/abs/1810.02338
19. Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B. Tenenbaum, Jiajun Wu. "The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision". ICLR 2019; arXiv:1904.12584. https://arxiv.org/abs/1904.12584
20. Trieu H. Trinh, Yuhuai Wu, Quoc V. Le, He He, Thang Luong. "Solving olympiad geometry without human demonstrations". Nature 625:476-482, 2024. https://doi.org/10.1038/s41586-023-06747-5
21. Trieu Trinh, Thang Luong. "AlphaGeometry: An Olympiad-level AI system for geometry". Google DeepMind blog, 17 January 2024. https://deepmind.google/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/
22. AlphaProof and AlphaGeometry teams. "AI achieves silver-medal standard solving International Mathematical Olympiad problems". Google DeepMind blog, 25 July 2024. https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/
23. Yuri Chervonyi, Trieu H. Trinh, Miroslav Olšák, et al. "Gold-medalist Performance in Solving Olympiad Geometry with AlphaGeometry2". arXiv:2502.03544, 5 February 2025; published in Journal of Machine Learning Research 26(241), 2025. https://arxiv.org/abs/2502.03544
24. Thomas Hubert, Rishi Mehta, Laurent Sartran, et al. "Olympiad-level formal mathematical reasoning with reinforcement learning". Nature 651:607-613, published online 12 November 2025. https://doi.org/10.1038/s41586-025-09833-y
25. Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, Graham Neubig. "PAL: Program-aided Language Models". arXiv:2211.10435, 18 November 2022. https://arxiv.org/abs/2211.10435
26. Liangming Pan, Alon Albalak, Xinyi Wang, William Yang Wang. "Logic-LM: Empowering Large Language Models with Symbolic Solvers for Faithful Logical Reasoning". Findings of EMNLP 2023; arXiv:2305.12295. https://arxiv.org/abs/2305.12295
27. Agnese Chiatti, Michael Cochez, Cristina Cornelio, et al. "The RAIL Principles for Neurosymbolic AI: Reasoning, Assurances, Interfacing and Learning". arXiv:2608.04285, 4 August 2026. https://arxiv.org/abs/2608.04285
28. Defense Advanced Research Projects Agency. "Assured Neuro Symbolic Learning and Reasoning (ANSR)". DARPA, accessed 18 November 2026. https://www.darpa.mil/research/programs/assured-neuro-symbolic-learning-and-reasoning
29. Werner Vogels. "A chat with Byron Cook on automated reasoning and trust in AI systems". All Things Distributed, 17 February 2026 (interview originally published in The Kernel). https://allthingsdistributed.com/2026/02/a-chat-with-byron-cook-on-automated-reasoning-and-trust-in-ai-systems.html
30. Rich Sutton. "The Bitter Lesson". incompleteideas.net, 13 March 2019. http://www.incompleteideas.net/IncIdeas/BitterLesson.html
31. Guillaume Lample, François Charton. "Deep Learning for Symbolic Mathematics". arXiv:1912.01412, 2 December 2019. https://arxiv.org/abs/1912.01412
32. Neurosymbolic Artificial Intelligence. "Neurosymbolic Artificial Intelligence journal". IOS Press, accessed 18 November 2026. https://neurosymbolic-ai-journal.com/
33. logictensornetworks. "logictensornetworks: Deep Learning and Logical Reasoning from Data and Knowledge". GitHub. https://github.com/logictensornetworks/logictensornetworks
34. tommasocarraro. "LTNtorch: PyTorch implementation of Logic Tensor Networks, a Neural-Symbolic framework". GitHub. https://github.com/tommasocarraro/LTNtorch
35. Henry A. Kautz. "The third AI summer: AAAI Robert S. Engelmore Memorial Lecture". AI Magazine 43(1):105-125, 2022. https://doi.org/10.1002/aaai.12036
36. Thang Luong, Edward Lockhart. "Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad". Google DeepMind blog, 21 July 2025. https://deepmind.google/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad/

