Meta-Interpretive Induction in MIL

• Meta-interpretive induction is a framework for learning logic programs via higher-order templates that enable compact, recursive rule synthesis.
• It employs meta-rules to guide predicate invention, ensuring that induced hypotheses are both interpretable and generalizable.
• The approach integrates symbolic and neuro-symbolic methods, supporting efficient search and scalable program synthesis in complex domains.

Meta-interpretive induction is a machine learning paradigm at the intersection of inductive logic programming (ILP), higher-order predicate invention, and symbolic/neuro-symbolic learning. It generalizes standard ILP by learning logic programs not only from data and background knowledge, but under a higher-order template bias specified as meta-rules—second-order clauses over predicate variables. This approach enables compact and recursively-defined logic programs, supports predicate invention, and provides a tractable framework for synthesizing interpretable, generalizable models from complex, structured data. Meta-interpretive induction underlies the Meta-Interpretive Learning (MIL) class of algorithms and its variants, which have been applied to program synthesis, causal world modeling, neuro-symbolic reasoning, and scientific knowledge discovery.

1. Formal Framework and Definitions

Meta-interpretive induction seeks to find a hypothesis $H$—a set of definite clauses—given background knowledge $B$, positive and negative examples $E^+, E^-$, and a finite set of meta-rules $M$ (higher-order second-order Horn clause templates). Formally, the MIL problem is to find $H \subseteq \text{Inst}(M)$ such that:

• $B \cup H \models e$ for all $e \in E^+$
• $B \cup H \not\models e$ for all $e \in E^-$
• $H$ is minimal or otherwise optimal with respect to a bias (e.g., smallest $|H|$ or minimal cost) (Patsantzis et al., 2021, Dai et al., 2021)

Each meta-rule $\mu \in M$ has the generic form: $$P(\mathbf{X}) \leftarrow Q_1(\mathbf{Y}_1), Q_2(\mathbf{Y}_2), ..., Q_k(\mathbf{Y}_k)$$ where $P, Q_i$ are predicate meta-variables. A hypothesis is a set of definite clauses built by instantiating meta-variables with predicate symbols from $B$ or invented predicates, and object variables with concrete terms. Predicate invention arises naturally as new predicate symbols instantiated into $P, Q_i$ without prior definition in $B$.

The induced hypothesis space is thus the class $$\mathcal{H}_{B,M} = \{ H \mid H \text{ is a set of ground instances of meta-rules in } M \}$$, restricted by $B$ and examples.

2. Algorithmic Realizations and Search Strategies

A core challenge is combinatorial explosion in the space of potential groundings. Algorithmic solutions in the literature exploit several techniques:

• Top Program Construction: The “Top program” is the complete set of ground meta-rule instantiations whose bodies are (potentially) satisfiable in $B$, constructed once in polynomial time under bounded arity and term depth (Patsantzis et al., 2021). This forms a finite hypothesis lattice, supporting efficient hypothesis selection and reduction.
• Specialization by Abduction: Upon encountering false negatives (missed positive examples) or false positives during verification, abductive reasoning is used to generalize or specialize the hypothesis by inventing new predicates and refining clause bodies (Crespo-Fernandez et al., 19 Feb 2026).
• Type-Directed Pruning: Introducing polymorphic and refinement types to $B$ and $M$ constrains instantiations to type-consistent ones, yielding up to a cubic reduction in search space as typed predicate slots restrict admissible symbol combinations. SMT-based refinement schemes further prune hypotheses by logical constraint satisfaction (Morel, 2021).
• Hybrid Symbolic/Neuro-Symbolic Schemes: MIL is integrated in an alternating optimization framework with neural perception modules (e.g., $P_\theta(z|x)$) and symbolic proof search. Abduction assigns pseudo-labels consistent with MIL hypotheses, back-propagating losses through perception and symbolic layers (Dai et al., 2020, Dai et al., 2021).
• Declarative Encodings: MIL can be encoded in Answer Set Programming with external sources (HEX), outsourcing background BK to mitigate grounding blowup and leveraging ASP's conflict propagation to prune the search (Kaminski et al., 2018).

The standard Metagol implementation realizes a query-driven, Prolog-based MIL, while the Louise and TOIL systems extend this with polynomial-time Top program construction and automated metarule learning (Patsantzis et al., 2021, Patsantzis et al., 2021).

3. Predicate Invention, Metarule Templates, and Expressivity

A hallmark of meta-interpretive induction is predicate invention—the dynamic introduction of new symbolic abstractions as reusable higher-level concepts or latent relations. Metarules determine the expressivity and generality of the hypothesis space. Common metarule templates include (Crespo-Fernandez et al., 19 Feb 2026, Dai et al., 2021, Dai et al., 2020):

• Chain: $P(X,Y) \leftarrow Q(X,Z), R(Z,Y)$
• Identity: $P(X,Y) \leftarrow Q(X,Y)$
• Recursion: $P(X,Y) \leftarrow Q(X,Z), P(Z,Y)$
• Absorption: $P(X,Y) \leftarrow Q(X,Y), R(Y)$

Metarule specialization organizes the hypothesis space as a lattice. The cardinality of the metarule language is polynomial in the number of literals in higher-order templates under suitable bounding conditions (Patsantzis et al., 2021). Invented predicates are registered canonically via hashing and memoization to guarantee sharing and eliminate redundant inventions (Crespo-Fernandez et al., 19 Feb 2026).

Metarule sets can themselves be learned (“metarule specialization”), increasing the flexibility and automating the bias previously chosen by the user (Patsantzis et al., 2021).

4. Integration With Abduction and Raw Data: Abductive Meta-Interpretive Learning

Abductive Meta-Interpretive Learning (Meta_Abd) unites abduction—the inference of latent symbolic facts explaining observed raw data—and induction of logic programs under the MIL paradigm (Dai et al., 2020, Dai et al., 2021). The formal generative model is: $$P(y, z, H | B, x, \theta) = P(y \mid B, H, z) \cdot P(H|B) \cdot P_\theta(z|x)$$ where $x$ are raw (sub-symbolic) inputs (e.g., images), $z$ are symbolic pseudo-labels abduced from $x$, and $H$ is the logic program induced by MIL.

Learning proceeds by alternating:

• Abductively inferring $z$ for $x$ under current $H$, constraining $z$ to be provable in $B \cup H \cup \{z\} \models y$;
• Updating perception model parameters $\theta$ by maximizing $P_\theta(z|x)$;
• Inducing or refining $H$ under a meta-rule grammar, possibly using a Bayesian prior on rule complexity.

This joint scheme significantly reduces combinatorial complexity versus naïve joint EM by interleaving abduction with induction, and allows recursive, interpretable program induction from raw perceptual data (Dai et al., 2020, Dai et al., 2021).

5. Practical Implementations and Empirical Performance

Empirical results across domains demonstrate that meta-interpretive induction yields human-interpretable, generalizable programs with superior data efficiency and scalability:

• Relational Causal World Modeling: In continual online learning of symbolic world models and lifted causal dynamics, MIL with dynamic predicate invention achieves zero-shot transfer across domain scales, rapid convergence, and sample efficiency orders of magnitude higher than neural PPO baselines (Crespo-Fernandez et al., 19 Feb 2026).
• Neuro-Symbolic Tasks: On MNIST-based reasoning (cumulative sum/product, sorting), Meta_Abd attains 95–97% symbolic accuracy and generalizes monotonically to inputs of greater complexity, outperforming LSTM/RNN and DeepProbLog baselines by wide margins in both predictive accuracy and computational cost (Dai et al., 2020).
• Synthetic Biology Knowledge Discovery: In automated biodesign, Meta_Abd learns compact models predicting protein yields with minimized experimental budgets, outperforming both symbolic and parametrized-only baselines (Dai et al., 2021).
• Program Synthesis and Language Semantics: MIL, extended to higher-order meta-rules, efficiently infers small-step operational semantics including predicate invention over function symbols, with empirical tractability via sequential induction and depth-bounded search (Bartha et al., 2019, Morel, 2021).
• Answer Set Programming Integration: HEX-formalism encodings of MIL combine Prolog's procedural bias with efficient ASP search, scaling to planning-heavy domains unapproachable by standard Prolog-based MIL (Kaminski et al., 2018).

Representative computational comparison (selected from (Patsantzis et al., 2021, Dai et al., 2020, Dai et al., 2021)):

System	Domain	Accuracy	Time (s)	Data Efficiency
Louise (MIL)	Grid 6x6	100%	1.78	—
Metagol	Grid 6x6	100%	12.4	—
Meta_Abd	MNIST cum. sum	95.3%	—	40 runs (vs. 100)
LSTM+NALU	MNIST cum. sum	0%	—	100 runs

MIL-based systems exhibit notable polynomial-time construction and hypothesis reduction, and their performance advantage widens with hypothesis space size or noisy/complex domains (Patsantzis et al., 2021, Dai et al., 2020).

6. Extensions: Types, Refinement, and Declarative Encodings

Recent work incorporates polymorphic Hindley–Milner style and refinement types, checked via SMT solvers, into MIL (Morel, 2021). Typed MIL restricts predicate instantiation at each meta-level, yielding a cubic asymptotic reduction in space. Refinement types—logical constraints on predicate arguments—enable further pruning through satisfiability tests, enhancing soundness and proof strength during clause construction. Type-directed MIL retains completeness for program classes expressible by the original meta-logic and substantiates empirically the predicted computational gains.

Declarative encodings (ASP/HEX) allow outsourcing BK, procedural bias recovery via relevance constraints, and overcoming Prolog's inherent limitations in handling grounding and conflict propagation (Kaminski et al., 2018).

7. Limitations, Scalability, and Theoretical Context

Despite polynomial-time construction of the Top program and type-based pruning, the overall complexity of hypothesis selection in large or unbounded arity settings can remain prohibitive; classic no-free-lunch results apply unless inductive bias is made explicit (Wolpert, 2021). Scalability to highly nondeterministic or cyclic domains may require additional planning heuristics or state abstraction mechanisms (Kaminski et al., 2018).

In summary, meta-interpretive induction and its MIL realizations constitute a powerful and theoretically grounded framework for interpretable program synthesis, lifted causal modeling, neuro-symbolic integration, and domain-aware scientific discovery. The approach combines the strengths of ILP, higher-order bias, predicate invention, and inductive abduction, underpinned by theoretical guarantees in terms of soundness, completeness, and complexity when adopting bounded hypothesis spaces and type systems.