7B Model Outsmarts Bigger LMs in Lean Proofs

A 7B parameter model outshines giants in Lean proof optimization.

ImProver 2 lays out a compelling case that neurosymbolic AI can turn small models into serious proof engines. The paper introduces a Lean 4 oriented framework that blends data efficient expert iteration with a scaffold that couples formal proof structure to lightweight informal abstractions. The result is a 7B parameter model that, within its family, outperforms orders of magnitude larger models and holds its own against mid tier frontier systems across a range of metrics. Crucially, the team shows the scaffold itself is not an afterthought, it meaningfully lifts performance for both small and larger models, suggesting proof optimization can be learned and scaled rather than hunted purely by brute force.

To put the numbers in perspective, the reported 7B model outperforms much larger peers in the same family and remains competitive with mid tier frontier models. That combination, strong performance at a fraction of the parameter count, plus cross model gains from the scaffolding, signals a potentially practical route for teams building formal verification assistants, refactoring tools for large libraries, or automated proof assistants that need to scale without extremely large compute budgets.

Analysts describe the outcome as a practical throughput upgrade for formal mathematics tooling. The paper demonstrates that properly scaffolded small models can reorganize “research level proofs” across varied metrics and do so with training dynamics that are more tractable than chasing ever larger architectures. In other words, the breakthrough isn’t simply a bigger brain; it’s a smarter brain with a map.

Industry takeaways and practitioner insights

Data efficiency matters, but it is not a free pass. ImProver 2 shows you can pull leverage from smaller models when you attach a scaffold that exposes formal structure and guides learning. Expect to balance scaffold design against the cost of training and inference in production settings.

The scaffold is the real lever. The performance boost for both small and frontier models hinges on the neurosymbolic scaffold plus the expert iteration loop. Without it, the gains from the same size model fade.

Expect brittleness in real world libraries. Formal proof work shifts as libraries grow and change; performance depends on stable formalization and robust evaluation metrics. Systems will need continuous calibration and potentially hooks into Lean 4's evolving tooling.

Generalization awaits broader proof systems. While results look promising in Lean 4, product teams should watch how well the approach generalizes to other theorem provers, libraries, and domain specific formalizations before bets scale company wide.

What this means for products shipping this quarter

Tooling for Lean and formal proof workflows could start shipping tighter, more capable proof suggestion and refactoring aids built on small, scaffolded models. Early adopters may pilot internal proof automation assistants that propose structured rewrites and steps, with users retaining final editorial control. The takeaway for startups is clear: invest in scaffolds and expert iteration loops now, and you may offer proof optimization workflows that are noticeably more cost effective than chasing ever larger models.

Sources

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

7B Model Outsmarts Bigger LMs in Lean Proofs

The Robotics Briefing