IneqAgent: Agentic Inequality Proving in Lean with Symbolic Computation

TEXT START: Formal inequality proving requires mathematical reasoning and exact computation, often connected through intermediate lemmas and goal reformulations.

The Dissection

The paper packages a narrow but consequential form of mathematical labor into an automated pipeline: an LLM proposes structure, symbolic engines generate exact certificates, Monte Carlo Tree Search selects lemma-level moves, and Lean verifies the result. Its real achievement is not “AI understands mathematics” in the grandiose sense. It is the industrialization of a previously expensive proof workflow.

The 94.09% benchmark result matters because the system does not merely generate plausible prose; it produces machine-checked proofs against original theorem statements. The labor unit being attacked is therefore substantive: lemma construction, reformulation, backend selection, and proof assembly.

The Core Fallacy

The text’s implicit conceptual error is treating benchmark success as the endpoint rather than evidence of labor displacement. It measures whether proofs close, not how much human mathematical production remains necessary around the system: theorem selection, formalization, benchmark design, debugging, maintenance, novelty assessment, and interpretation.

Under the Discontinuity Thesis, this distinction is not comforting. If agentic systems can absorb the proof-search loop while symbolic tools supply exact certificates and Lean supplies verification, the human mathematician is no longer the exclusive producer of economically valuable proof labor. The system may not yet satisfy universal Cognitive Automation Dominance, but it directly reinforces P1 in a domain often treated as resistant to automation.

The paper also risks confusing formal verification with human indispensability. Lean eliminates certain classes of error; it does not preserve the economic necessity of the person who formerly generated the proof.

Hidden Assumptions

• The four benchmarks represent real-world inequality proving rather than a favorable slice of the task distribution.
• The reported success rate generalizes beyond the selected theorem families and available symbolic backends.
• Low inference cost remains low when theorem formalization, environment setup, failed attempts, debugging, and maintenance are included.
• Human experts are not still supplying the decisive decomposition, conjecture selection, or backend engineering.
• Theorem statements are already formalized in Lean, despite formalization itself being a major labor bottleneck.
• Success on known conjectures transfers to genuinely novel mathematics rather than sophisticated search over recognizable structures.
• Lean verification proves the intended theorem is useful, not merely that the encoded statement is correct.
• The evaluated backbone and toolchain remain competitive as models and symbolic systems evolve.
• More powerful proof agents will remain complements to mathematicians instead of becoming substitutes for large portions of their workflow.
• The market will continue paying for proof production as a distinct human service once verified proof generation becomes cheap and scalable.

Social Function

Primary classification: partial truth and transition management, with a layer of prestige signaling.

The partial truth is real: exact proof construction is being automated, and formal verification makes the result materially stronger than ordinary language-model output. The transition-management function lies in presenting this as a framework and benchmark advance rather than foregrounding the labor-market consequence: proof production is being decomposed into searchable, certifiable, machine-executable operations.

The prestige layer is the leaderboard framing—backbones, success rates, inference costs, and named conjectures. Those metrics are useful technically, but they can make a structural labor shift look like an incremental research competition. The paper is not mere copium. It is a clean demonstration of a mechanism that makes copium harder to maintain.

The Verdict

IneqAgent is a narrow but credible breach in the wall protecting high-status cognitive labor. It does not, from the supplied evidence alone, prove the full Discontinuity Thesis or establish universal AI dominance. It does prove something more immediate: a specialized class of expert reasoning can be converted into an agentic search process, exact symbolic subroutines, and mechanically verified output.

The human role is already being pushed upward—from proving to specifying, supervising, formalizing, and judging significance. Those are temporary moats, not permanent exemptions. If the same pattern scales across adjacent mathematical and technical domains, the proof-producing majority becomes economically ornamental while ownership of the agentic infrastructure captures the value. The paper is therefore a technical result with a larger implication: formal mathematics is not outside the automation frontier; it is being turned into one of its cleanest laboratories.