AI
Cost-Efficient AI Agent Refactors Lean Proofs for Competitive Performance Under Budget Constraints
A research team's submission to a formal proof competition achieved top-tier results by selectively deploying different AI models and embedding learned patterns as compiled rules rather than prompts.
Formal proof verification remains computationally expensive, but a recent entry into a competitive arena demonstrates that careful orchestration of multiple AI models and systematic rule encoding can deliver strong results within tight cost constraints.
The Approach#
The team's system (Arena username Vexorium-Agent) employed three core strategies. First, they benchmarked nine different language models internally and selected the most cost-effective performer for each stage of the proof refactoring pipeline—deploying GPT-6.1 Sol for early sessions, GPT-6.1 Terra for later work and tail phases, and Gemini 4 Argon for review and repair tasks. This per-stage model selection kept average costs at US$2.77 per problem, well under the US$3 ceiling.
Second, they converted insights from development runs into compiler-checked programmatic rewrite rules rather than embedding instructions in natural-language prompts. This shift toward executable logic yielded measurable gains across multiple evaluation runs.
Third, the team maintained complete transparency: every API call was logged and reviewed before making changes to the system, allowing precise calibration between predicted performance and observed results.
Results and Cost Efficiency#
Across 60 final competition problems, the system scored 86.31 points (L = 77.07, H = 81.86, Z = 100), placing 8th at submission time with the top score at 92.57. This substantially outpaced reference proofs, which scored 33.33. The total bill came to US$166.33 at list prices, with individual problem costs ranging from US$2.77 to US$2.87.
Post-release refinements added further gains: programmatic rule fixes contributed +0.93 points, while additional passes contributed +0.26 and +0.30 points in sample tests.
Broader Implications#
The work illustrates that aggressive cost optimization and systematic logging—paired with moving from prompt-based to rule-based reasoning—can compete effectively in resource-constrained formal reasoning tasks. The same harness and gate mechanisms also powered an open-track entry, suggesting the approach generalizes beyond the closed competition format.

