ImProver 2: Iteratively Self-Improving LMs for Neurosymbolic Proof Optimization

URL SCAN: ImProver 2: Iteratively Self-Improving LMs for Neurosymbolic Proof Optimization

FIRST LINE: Formal mathematics libraries are rapidly expanding, creating a growing need to refactor verified proofs for maintainability and to improve training data quality for neural provers.

The Dissection

This paper documents a self-referential bootstrap in AI capability: a model that optimizes the proofs used to train future models. The headline 7B parameter model outperforming "orders-of-magnitude larger models" is the scaffolding story — better architecture of the reasoning process beats brute force scaling. The real signal is buried in the functional description: proof optimization is now a learnable, scalable task that smaller systems can execute at research-level quality.

The Core Fallacy (DT Frame)

The paper presents this as a technical contribution to formal mathematics tooling. It is. But from the DT lens, this is evidence of P1 tightening. Mathematical proof generation has been the canonical last refuge of "pure human reasoning" — the domain where AI supposedly struggles with genuine novelty and deep abstraction. ImProver 2's results suggest:

1. The cognitive work is becoming mechanizable — not just proof search, but proof restructuring (refactoring for maintainability, which requires metajudgment about quality and structure, not just completion).
2. The bootstrapping loop is closing — AI optimizing AI training data quality is a self-improving loop that accelerates the very systems designed to render mass human cognitive labor redundant.
3. Scaling is being decoupled from capability — the fact that a 7B model matches much larger ones via scaffolding means the capability ceiling is being reached through architectural efficiency, not raw parameter count. This widens the threat surface dramatically.

The Hidden Assumption

The paper assumes formal mathematics is a relatively contained domain where this approach can be benchmarked cleanly. But this containment is eroding. The neurosymbolic scaffold — formal structure + informal abstraction — mirrors exactly the kind of hybrid reasoning that transfers to domain-adjacent cognitive work: engineering design, scientific hypothesis generation, code synthesis, economic modeling. The scaffolding lesson learned here scales.

Social Function

Prestige signaling for the scaling-is-dead thesis — this paper is a grenade thrown at pure LLM scaling advocates. If efficient scaffolding + training pipeline optimization beats raw parameter count, the race shifts to who controls the proof optimization infrastructure, not who builds the biggest model. This is both technically interesting and strategically significant: the moats in AI capability are shifting from model weight to data quality and reasoning architecture.

The Verdict

This is a functional demonstration of accelerating cognitive automation wrapped in formal methods theater. The proof optimization loop — AI improving the training data that trains AI — is exactly the bootstrap mechanism the DT timeline depends on. Every paper like this is a data point confirming that P1 is not theoretical, it's being engineered in real time across cognitive domains simultaneously. The lag between "AI can do X in a lab" and "AI does X at scale, rendering human labor in that domain non-survival" is compressing. ImProver 2 doesn't just optimize proofs. It optimizes the path to post-WWII capitalism's structural obsolescence.