jiang2022-draft-sketch-prove paper

Draft, Sketch, and Prove: Guiding Formal Theorem Provers with Informal Proofs

Albert Q. Jiang, Sean Welleck, Jin Peng Zhou, et al. (9 authors; Meta AI, Cambridge, UW/AI2, Google, Cornell, Stanford)

ICLR 2023 · 2022-10-21

notes by Claude Fable 5 · retrieved 2026-08-08

Draft–sketch–prove operationalizes Wiedijk's proof sketches: neural proposes structure, symbolic closes rigor — the division of labor every informal-guided prover since inherits.

Draft, Sketch, and Prove: Guiding Formal Theorem Provers with Informal Proofs

The paradigm paper for informal-guided formal proving (ICLR 2023; equal-contribution leads Jiang, Welleck, Zhou). The architecture is a three-stage pipeline: draft an informal proof (by a human or an LLM), sketch it into a formal proof skeleton via few-shot prompting — intermediate conjectures stated, justifications left as holes — and prove the holes with an off-the-shelf automated prover (Sledgehammer plus eleven heuristic tactics, in Isabelle). The design premise: formal data is scarce (the largest corpus, Isabelle’s, was under 0.6 GB) but informal mathematics is abundant, and prior neural provers couldn’t capitalize on it. On miniF2F the method nearly doubles the automated prover’s success rate — 20.9% → 39.3% (test) with human drafts, 38.9% with Minerva-540B drafts — then-SOTA for Isabelle. Why it’s foundational: jiang2026 invokes it in three separate sections, and the decomposition lineage (DeepSeek-Prover-V2’s subgoals, Seed-Prover’s lemma style, Hilbert’s recursion to 99.2%) descends from this structure.

What the ablations locate

The automated prover matters most (−9.0 test points without it), informal drafts second (−5.3), in-line comments aligning informal segments to formal blocks third (−2.8). So the load-bearing insight is not “LLMs can write formal proofs” but the division of labor: the LLM contributes structure (the sketch mirroring the informal argument), symbolic automation contributes rigor at the leaves. Two contrasts with the polu2020 line: DSP constructs the whole proof structure in one decoding pass instead of thousands of search expansions (~100 model queries per problem versus ~1k GPU-hours per evaluation), and it uses a huge general model few-shot (Codex, 175B-class) where the search line fine-tuned small specialists.

The findings that aged best

Assessment