# The Limits of LLM Mathematical Theorem Proving and Verifiable Evaluation
Large language models can now solve competition math, but scientific research is a far harder benchmark. This article examines verifiable evaluation of LLM mathematical theorem proving and construction tasks: why machine-checkable verification matters, what recent evaluation results reveal about the capability gap between leading and second-tier models, and where AI research abilities hit their limits. It includes concrete results such as a 124-vertex, 630-edge extremal graph construction and a tightened coding-theory upper bound of A(22,10) ≤ 82, plus practical guidance for developers on building AI workflows around verifiability.
1. Why Verifiable Research Evaluation Is Needed
Competition math problems have a unique answer that can be checked immediately. Research tasks are different: they require the researcher to pose a question, explore approaches, construct proofs, and pass peer review. Evaluations that only look at final answers have an obvious flaw: a model may produce a proof that looks plausible but falls apart under scrutiny, while the accuracy number hides the problem.
To address this, researchers have rewritten classic problems from theoretical computer science into task sets that can be verified by machines, so that every step of a model’s output can be independently checked. This design rests on three principles.
- Deterministic verification: each problem comes with a formal or programmable judge. For example, in a graph construction task, the checker verifies whether the graph contains triangles or 4-cycles and counts edges to score the result.
- Traceable process: the model must output intermediate lemmas, construction parameters, or proof scripts that compile and run — not just a conclusion.
- Distinguish capability tiers: tasks are divided into three levels — producing a correct proof, improving a known construction, and tightening a known upper bound. The tiering prevents a single aggregate score from masking capability differences.
2. What the Evaluation Results Show
2.1 The Overall Picture: Few Leaders, a Wide Gap
In the overall results, the best-performing model clearly leads, a second tier follows behind, and the remaining agents score consistently low. Only a small number of models are capable of handling scientific reasoning, and the capability divide is pronounced.
In the theorem-proving subtask, the leading model solved roughly 12 of 14 problems — about 86% — without intermediate lemma hints, while second-tier models performed noticeably weaker. The original evaluation did not state the number of attempts or statistical methodology, so these figures should be read as relative comparisons rather than absolute capability measures.
Notably, one second-tier model completed a highly difficult problem (called the PCP alphabet soundness problem in the original report) on every attempt, while the top-ranked model failed to solve it. This suggests that different models have complementary reasoning strengths, and no single ranking can fully capture their capabilities.
2.2 The Most Convincing Evidence Comes from Construction Tasks
Compared with multiple-choice questions and proof-writing, extremal graph construction better reflects a model’s creative research ability. Two representative results stand out.
- Erdős extremal graph of girth 5: the task requires constructing a graph containing neither triangles nor 4-cycles. According to the evaluation record, the model produced a concrete construction with 124 vertices and 630 edges, beating the reference value of 629 edges used by the task. This construction can be independently verified — it is a genuinely new construction, not a reproduction of a known answer.
- Tightening a coding theory upper bound: A(n,d) denotes the maximum number of codewords achievable by a binary code of length n and minimum distance d. The evaluation record shows the model proved A(22,10) ≤ 82, tightening the upper bound of 84 in the existing literature to 82. The improvement is only a constant of two, but constant-factor improvements usually require intricate combinatorial arguments — and because the result is verifiable, it has genuine research value.
3. Capability Boundaries: What AI Can and Cannot Do
Taken together, these results allow AI research capabilities to be grouped into three tiers:
| Capability tier | What AI can do | What AI cannot yet do | How it is verified |
|---|---|---|---|
| Controlled reasoning | With a well-defined problem, sufficient budget, and per-step feedback: write proofs, optimize programs, improve constructions | Reach correct conclusions reliably without feedback | Checkers, compiling and running, formal verification |
| Autonomous research | Large-scale exploration and trial-and-error under a given goal | Autonomously select promising directions and judge the significance of results | Human review and peer evaluation |
| Open problems | Help with incremental progress within existing frameworks | Establish new frameworks or open new research directions | No reliable automated judgment exists |
One phenomenon in the second tier deserves attention. After long stretches of failure, models often struggle to switch approaches the way human researchers do. A plausible explanation is that the bottleneck lies not only in single-step reasoning but also in self-correction and strategy adjustment — though this needs further experimental validation.
4. Practical Implications
For developers and researchers already working with AI, several practices can be adopted directly:
- Prioritize verifiable tasks: code review, unit test generation, formal proof assistance, and parameter tuning all come with built-in verifiers. The value of outputs can be quickly confirmed or rejected, minimizing risk and maximizing benefit.
- Pair unverifiable long-form output with checks: design documents and architecture arguments are prone to high-confidence errors. Combine them with tests, review checklists, or executable prototypes; otherwise mistakes are hard to catch.
- Adopt a human-AI division of labor: let the model explore at scale under a clearly defined objective, while humans define the problem, set verification criteria, and judge the value of the results.
To apply this paradigm to test case generation, code refactoring, or performance tuning, first check the following:
- Can the objective be described in one sentence, with a clear success criterion?
- Is there an automated judge, and has the judge itself been validated?
- Is the intermediate process traceable — for example, saving intermediate results, parameters, and run logs?
- Are compute budget and attempt counts fixed in advance, and are results reproducible?
- Who reviews the output, and how are review results recorded?
On the infrastructure side, verification tasks require isolated execution environments, auditable logs, appropriate alerting, and clearly defined permission boundaries. These conditions make the exploration process reproducible and traceable after the fact.
5. Conclusion
The value of this evaluation lies in mapping the boundary of AI research capability. When the problem is well-defined, the budget sufficient, and the feedback verifiable, models can already produce auditable proofs and constructions that exceed established baselines. Choosing research directions, judging the significance of results, and tackling major open problems remain largely human work. For developers, the pragmatic move is to make verifiability a core principle of AI workflows — it is both what made this experiment succeed and the foundation of future human-machine collaboration.
