Pairing statistical LLM candidate beam generation with strict deterministic proof kernel verification
Large Language Models (transformers) operate on statistical token transitions. While remarkably proficient at proposing creative lemma substitutions and guiding proof search heuristics through vast combinatorial spaces, their lack of an internal truth model causes subtle mathematical hallucinations (e.g., falsely assuming non-commutative group elements commute or applying false geometric congruence).
Scores candidate next steps via learned mathematical priors. Discovers non-obvious auxiliary constructions (like in Google DeepMind's AlphaGeometry) that purely brute-force symbolic search could take centuries to encounter.
Executes strict axiomatic unification (Lean, Coq, Isabelle). Instantly rejects invalid leaps with zero tolerance for probabilistic "vibes", ensuring generated proofs meet rigorous mathematical certainty.