Neurosymbolic Learning and Reasoning

Integration targets and historical foundations

Neural–symbolic integration — 2021. Combines learned representations with symbolic knowledge or computation. Sarker et al. organize integration approaches; injecting a rule into training differs from executing a symbolic program at inference.
KBANN — 1994. Towell and Shavlik translate propositional domain theories into a neural network, then refine it by backpropagation. Rules initialize the architecture; subsequent training need not preserve them as hard constraints.
Abductive learning — 2019. Alternates learned perception with logical corrections of perceived facts. Dai et al. jointly learn digit recognition and mathematical operations from handwritten equations; logical consistency supplies training feedback for perception.

Logic Tensor Networks — 2020. Represent first-order knowledge in a many-valued differentiable logic and optimize satisfaction alongside learning. Truth degrees depend on the chosen logical operators; they are not automatically calibrated probabilities.
Logical Neural Networks — 2020. Associate neurons with components of logical formulas and propagate bounds on truth values. Contradiction-aware learning and an open-world treatment support incomplete knowledge within the proposed logic.
Semantic loss — 2018. Penalizes probability mass assigned to outputs that violate a propositional constraint. Xu et al. connect neural predictions to weighted model counting; a finite soft penalty differs from enforcing validity at decoding.
SATNet — 2019. Differentiates through a semidefinite relaxation of MAXSAT to integrate constraint solving with neural perception. Wang et al. test parity and Sudoku; the relaxed solver is not a general exact SAT oracle.
NeuroSAT — 2018. A message-passing network learns from SAT/UNSAT labels; satisfying assignments can emerge in its activations. More iterations solve harder tested instances, but the authors report lower reliability than conventional SAT solvers.

Knowledge compilation — 2002. Darwiche and Marquis compare logical representations by succinctness and tractable queries or transformations. Compiling a theory can make repeated reasoning cheap while enlarging the representation; tractability is query- and language-specific.

DeepProbLog — 2018. Adds neural predicates to probabilistic logic programs and trains neural and symbolic components jointly from examples. Program structure supplies explicit reasoning; perceptual uncertainty enters through neural probabilities.
DeepStochLog — 2021. Neural grammar rules define probabilities over derivations in stochastic logic programs. DeepProbLog instead uses possible-world semantics; the choice changes what a probability represents and how inference is computed.
DeepProofLog — 2025; AAAI 2026. Learns neural guidance for derivation steps in stochastic logic programs. A mapping from resolution to Markov decision processes connects proof search with dynamic programming and reinforcement learning.
NeurASP — 2020. Treats neural outputs as distributions over atomic facts in answer-set programs. Rules can constrain perception and training; the paper reports substantial symbolic-reasoning overhead relative to neural training alone.
Scallop — 2023. Combines Datalog-style rules, recursion, aggregation and negation with differentiable provenance-based reasoning. It offers a programming interface for integrating neural inputs with structured computations.

Differentiable ILP — 2017. Evans and Grefenstette learn explanatory rules through a differentiable inductive-logic framework. Experiments address noisy symbolic data and coupling to neural perception; the available rule language bounds expressivity.
Neural theorem proving — 2017. Rocktäschel and Riedel replace exact symbolic unification with similarity between learned embeddings inside backward chaining. This permits gradient learning of rules and symbols while changing the semantics of a proof score.
NeuralLP — 2017. Learns compositions of differentiable logical operations for knowledge-base reasoning. A neural controller jointly selects rule structure and weights; results concern relational queries rather than unrestricted theorem proving.
DreamCoder — 2020. Alternates program search, abstraction discovery and training a neural search guide on replayed or generated tasks. Learned libraries reuse symbolic components across tasks; the supplied programming language shapes the search.

Neuro-Symbolic Concept Learner — 2019. Learns object representations and question-to-program parsing from image/question/answer supervision. Symbolic programs execute over learned scene concepts; correct answers alone do not establish correct concept grounding.
AlphaGeometry — 2024. A neural language model proposes constructions for a symbolic geometry engine. Trinh et al. solve 25 of 30 selected olympiad problems using synthetic training data; the result concerns Euclidean plane geometry.
AlphaGeometry2 — 2025. Extends the geometry language, symbolic engine and search-tree knowledge sharing. The authors report solving 84% of IMO geometry problems from 2000–2024; this expands both language coverage and search beyond AlphaGeometry.
LeanDojo / ReProver — 2023. Retrieves premises for language-model-guided proof search in Lean. Its novel-premise split tests proofs using lemmas absent from training proofs; checked formal proofs provide a different output guarantee from soft neural proof scores.

Reasoning shortcuts — 2024. rsbench tests whether correct downstream answers hide incorrect learned concepts. Bortolotti et al. find concept quality remains difficult for neural and neurosymbolic models; task accuracy alone misses these failures.
Concept identifiability — 2023. Marconato et al. characterize reasoning shortcuts as unintended optima of the learning objective. Reconstruction and partial concept supervision have limits: matching task labels and rules can leave multiple incompatible concept meanings.
Separate perception from reasoning. Measure concept accuracy, answer accuracy and constraint violations separately; test new compositions and inspect which rules are given. A correct symbolic derivation can still start from an incorrect perceptual fact.

DeepProbLog implementation. Official probabilistic-programming system with neural predicates and example tasks. The probabilistic program, neural architecture and inference configuration jointly define the experiment.
LTNtorch. PyTorch implementation of Logic Tensor Networks with logical operators, quantifiers and satisfaction objectives. Record the fuzzy operators and aggregation choices when comparing models.

Star Issue