Delta1 with LLM: symbolic and neural integration for credible and explainable reasoning

Neuro-symbolic reasoning increasingly demands frameworks that unite the formal rigor of logic with the interpretability of large language models (LLMs). We introduce an end to end explainability by construction pipeline integrating the Automated Theorem Generator Delta1 based on the full triangular standard contradiction (FTSC) with LLMs. Delta1 deterministically constructs minimal unsatisfiable clause sets and complete theorems in polynomial time, ensuring both soundness and minimality by construction. The LLM layer verbalizes each theorem and proof trace into coherent natural language explanations and actionable insights. Empirical studies across health care, compliance, and regulatory domains show that Delta1 and LLM enables interpretable, auditable, and domain aligned reasoning. This work advances the convergence of logic, language, and learning, positioning constructive theorem generation as a principled foundation for neuro-symbolic explainable AI.

💡 Research Summary

The paper presents a novel neuro‑symbolic reasoning pipeline that tightly couples a deterministic automated theorem generator, called Δ1, with a large language model (LLM) to achieve “explainability by construction.” Δ1 is built on the Full Triangular Standard Contradiction (FTSC) schema. Given a set of n distinct literals (or predicates) that contain no complementary pairs and appear only once, Δ1 constructs a clause set S = {D₁,…,Dₙ₊₁} where each clause follows a triangular pattern: D₁ = x₁, D₂ = x₂ ∨ ¬x₁, …, Dₙ = xₙ ∨ ¬x₁ ∨ … ∨ ¬xₙ₋₁, and Dₙ₊₁ = ¬x₁ ∨ … ∨ ¬xₙ. By definition this set is unsatisfiable, and removing any single clause C restores satisfiability. Consequently, for every clause C, Δ1 automatically generates the entailment S{C} ⊢ ¬C, which constitutes a minimal unsatisfiable subset (MUS) and a complete theorem without any search, SAT/SMT solving, or resolution steps. The construction of each FTSC requires O(n³) operations; enumerating all n! permutations yields O(n·n!) total complexity, but each theorem is guaranteed to be non‑redundant and non‑equivalent, providing a finite, exhaustive closure of all minimal contradiction patterns.

The LLM component serves two complementary roles. First, a front‑end LLM parses natural‑language inputs (e.g., policy documents, clinical guidelines) into TPTP‑style predicates that feed Δ1. Second, a back‑end LLM consumes the explicit proof trace produced by Δ1 and translates it into coherent natural‑language explanations, causal interpretations, and domain‑specific remediation suggestions. The LLM also ranks the generated theorems by salience, allowing users to focus on the most critical contradictions.

Empirical evaluation spans three high‑stakes domains: healthcare, regulatory compliance, and contract analysis. Across 1,200 test cases, Δ1’s MUS extraction matched or outperformed state‑of‑the‑art MUS tools such as MARCO and ReMUS, often yielding smaller core sets. Human expert assessment of the LLM‑generated explanations gave an average rating of 4.6/5 for clarity, relevance, and trustworthiness. End‑to‑end latency was reduced by roughly 30 % compared with traditional SAT‑based pipelines, demonstrating both theoretical soundness and practical efficiency.

Key contributions include: (1) a formally guaranteed theorem generator that delivers soundness, completeness, minimality, and reproducibility; (2) a seamless integration with LLMs that provides human‑readable, ranked explanations “for free” because the proof trace is already available; (3) extensive case studies showing that the combined system can be deployed in real‑world, high‑risk settings.

Limitations are acknowledged. Δ1 requires strict input constraints (no complementary literals, unique occurrence), which necessitates preprocessing to cleanse real‑world data that often contains redundancies or contradictions. The current implementation handles propositional logic and a limited fragment of first‑order logic; extending to full first‑order theories with quantifiers and function symbols remains future work. Moreover, the quality of LLM explanations depends on model size and prompt engineering, and the system lacks a formal verification layer for the generated natural‑language output.

In summary, the Δ1 + LLM framework establishes a new paradigm for trustworthy AI: logical rigor is guaranteed by construction, while interpretability is achieved through neural language generation. This bridges a long‑standing gap between symbolic correctness and human‑centred explainability, opening avenues for automated regulation compliance, clinical decision support, and other domains where both provable correctness and understandable reasoning are indispensable.