Rögnvaldsson, Sun, Dekoninck and Vechev (ETH Zurich), arXiv 2606.04883, ICML 2026; MathArena blog; code at eth-sri/optimizing-lean-agents.
The formal checker is used here as a budget signal, not just a pass/fail gate. Agentic provers spend fixed budgets per target, so most compute goes to lemmas that are false, misformalized or out of reach. This paper adds a control plane that decides, after each pair of attempts, whether to keep trying a lemma breakdown or discard it.
Mechanism. Data plane: GPT-OSS-20B writes a natural-language lemma breakdown, Goedel-Formalizer-V2-8B formalizes it, and Goedel-Prover-V2-8B attempts each lemma. Control plane: attempt only while estimated quality minus lambda times estimated cost stays positive. Quality is a logistic regression on three cheap features: how similar the failed proofs are, how diverse the Lean errors are, and the attempt count. Cost is the average output tokens so far. Error diversity is the strongest feature.
Results. On a held-out 43-problem slice of PutnamBench: 28.9% less compute (measured as effective parameters times output tokens) at equal solve rate, or +7.9% solve rate at equal compute. Example: 41.6% solved for 15.7M units versus 41.8% for 22.3M under fixed budgets. An oracle router would save 62 to 67% (paper and blog disagree), so most of the headroom is unclaimed. With 10% label noise the saving drops to 17%.
Why it matters for efficiency. Compiler and checker errors are free, exact and structured. Their diversity tells you whether an agent is making progress or circling, which is the question every retry loop needs answered.
Caveats. 85 problems total (42 train, 43 test); about 1,600 to 2,000 GH200-hours for the study. Only attempt-or-terminate actions, no model switching. One prover family.