SYNTHESIS NOTE
Topics›Natural Language Inference›this note

Can autoformalization work on individual statements alone?

Explores whether translating single theorems into formal logic is feasible without building entire supporting libraries of axioms, definitions, and lemmas first.

Synthesis note · 2026-07-17 · sourced from Natural Language Inference

Autoformalization translates informal natural language into machine-verifiable formal languages, and almost all of it operates statement-by-statement. This position paper argues that framing is a category error: real formalization is inherently theory-level. Before a target theorem like Pythagoras can even be stated, you need an entire "tower" beneath it — axiomatic primitives, derivative definitions, notations, examples, supporting lemmas, proof tactics, and all their inter-dependencies — assembled as a coherent formal library. Statement-level autoformalization only looks tractable because it silently borrows that scaffolding from a prebuilt context like Lean's Mathlib.

The evidence is the sheer cost of real efforts: formalizing the Kepler conjecture — a single statement — took 11 years, almost all spent constructing the web of definitions and lemmas needed to support it. Which means the interesting bottleneck is not translating one sentence faithfully but building and maintaining a self-consistent library of interdependent formal objects.

This sharpens the vault's autoformalization thread. Since Can large language models translate natural language to logic faithfully?, per-statement translation is already unreliable; theory-level formalization compounds that risk because dependency errors propagate — a mis-formalized definition silently corrupts every lemma above it, so semantic-loss failures no longer stay local. It also reframes the trade-off behind Can structured templates replace formal verification for code reasoning?: if full theory-level formalization is this expensive, structured-but-informal scaffolding is not a stopgap but a rational point on the cost curve for most reasoning tasks.

Inquiring lines that read this note 36

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

Can external verification replace human reasoning oversight for correctness? How can we verify and preserve mathematical understanding in AI-generated proofs? Do language models learn genuine understanding or just surface patterns? How do symbolic and semantic approaches combine in LLM reasoning? What drives the effectiveness and failure of language model reasoning? How does compressing or distilling information affect model learning and generalization? What causes models to fail in specialized domains versus generally? Does AI coding assistance improve productivity and code quality across experience levels? How does the generation-verification gap constrain AI self-improvement capabilities? Can AI research automation improve discovery pace without sacrificing quality? How are knowledge and reasoning functionally separated in neural network architecture?

Related concepts in this collection 3

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
12 direct connections · 127 in 2-hop network ·dense cluster Open in graph ↗

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Original note title

autoformalization must move from isolated statements to whole theories — a target theorem is unstatable without its web of axioms definitions and lemmas