We present Generative Logic (GL), a deterministic architecture that starts from user-supplied axiomatic definitions (and, optionally, a list of simple facts for counterexample (CE) construction), written in a minimalist Mathematical Programming Language (MPL), and systematically explores their deductive neighborhood. Definitions are compiled into a distributed grid of simple Logic Blocks (LBs) that exchange messages; whenever the premises of an inference rule unify, a new fact is emitted with full provenance to its sources, yielding replayable, auditable proof graphs. A prototype software implementation instantiates the workflow on first-order Peano arithmetic. Starting only from the Peano axioms, GL enumerates conjectures, applies normalization, type, and CE filter, and automatically reconstructs machine-checkable proofs of foundational arithmetic laws, including associativity and commutativity of addition, associativity and commutativity of multiplication, and distributivity. On commodity hardware, the prover phase requires approximately 7 seconds; a complete run finishes in about 5 minutes. Generated proofs export to navigable HTML so that every inference step can be inspected independently. We outline a hardware-software co-design path toward massively parallel realizations and describe prospective integration with probabilistic models (e.g., large language models) for auto-formalization and conjecture seeding. The Python, C++, and MPL code to reproduce the Peano experiments, along with the full proof graphs in HTML as well as machine-readable text format, are available in the project's GitHub repository at github.com/Generative-Logic/GL commit 56c9233 and are permanently archived at doi:10.5281/zenodo.17206386.

Reasoning over natural language is a challenging problem in NLP. In this work, we focus on proof generation: Given a hypothesis and a set of supporting facts, the model generates a proof tree indicating how to derive the hypothesis from supporting facts. Compared to generating the entire proof in one shot, stepwise generation can better exploit the compositionality and generalize to longer proofs but has achieved limited success on real-world data. Existing stepwise methods struggle to generate proof steps that are both logically valid and relevant to the hypothesis. Instead, they tend to hallucinate invalid steps given the hypothesis. In this paper, we present a novel stepwise method, NLProofS (Natural Language Proof Search), which learns to generate relevant steps conditioning on the hypothesis. At the core of our approach, we train an independent verifier to check the validity of the proof steps to prevent hallucination. Instead of generating steps greedily, we search for proofs maximizing a global proof score judged by the verifier. NLProofS achieves state-of-the-art performance on EntailmentBank and RuleTaker. Specifically, it improves the correctness of predicted proofs from 27.7% to 33.3% in the distractor setting of EntailmentBank, demonstrating the effectiveness of NLProofS in generating challenging human-authored proofs.

We study abduction in First Order Horn logic theories where all atoms can be abduced and we are looking for preferred solutions with respect to three objective functions: cardinality minimality, coherence, and weighted abduction. We represent this reasoning problem in Answer Set Programming (ASP), in order to obtain a flexible framework for experimenting with global constraints and objective functions, and to test the boundaries of what is possible with ASP. Realizing this problem in ASP is challenging as it requires value invention and equivalence between certain constants, because the Unique Names Assumption does not hold in general. To permit reasoning in cyclic theories, we formally describe fine-grained variations of limiting Skolemization. We identify term equivalence as a main instantiation bottleneck, and improve the efficiency of our approach with on-demand constraints that were used to eliminate the same bottleneck in state-of-the-art solvers. We evaluate our approach experimentally on the ACCEL benchmark for plan recognition in Natural Language Understanding. Our encodings are publicly available, modular, and our approach is more efficient than state-of-the-art solvers on the ACCEL benchmark.

What capabilities are required for an AI system to pass standard 4th Grade Science Tests? Previous work has examined the use of Markov Logic Networks (MLNs) to represent the requisite background knowledge and interpret test questions, but did not improve upon an information retrieval (IR) baseline. In this paper, we describe an alternative approach that operates at three levels of representation and reasoning: information retrieval, corpus statistics, and simple inference over a semi-automatically constructed knowledge base, to achieve substantially improved results. We evaluate the methods on six years of unseen, unedited exam questions from the NY Regents Science Exam (using only non-diagram, multiple choice questions), and show that our overall system’s score is 71.3%, an improvement of 23.8% (absolute) over the MLN-based method described in previous work. We conclude with a detailed analysis, illustrating the complementary strengths of each method in the ensemble. Our datasets are being released to enable further research.

In predicate invention (PI), new predicates are introduced into a logical theory, usually by rewriting a group of closely-related rules to use a common invented predicate as a "subroutine". PI is difficult, since a poorly-chosen invented predicate may lead to error cascades. Here we suggest a "soft" version of predicate invention: instead of explicitly creating new predicates, we implicitly group closely-related rules by using structured sparsity to regularize their parameters together. We show that soft PI, unlike hard PI, consistently improves over previous strong baselines for structure-learning on two large-scale tasks.

Alpha-Beta is the most common game tree search algorithm, due to its high-performance and straightforward implementation. In practice one must find the best trade-off between heuristic evaluation time and bringing the subset of nodes explored closer to a minimum proof graph. In this paper we present a series of structural properties of minimum proof graphs that help us to prove that finding such graphs is NP-hard for arbitrary DAG inputs, but can be done in linear time for trees. We then introduce the class of fastest-cut-first search heuristics that aim to approximate minimum proof graphs by sorting moves based on approximations of sub-DAG values and sizes. To explore how various aspects of the game tree (such as branching factor and distribution of move values) affect the performance of Alpha-Beta we introduce the class of ``Prefix Value Game Trees'' that allows us to label interior nodes with true minimax values on the fly without search. Using these trees we show that by explicitly attempting to approximate a minimum game tree we are able to achieve performance gains over Alpha-Beta with common extensions.