Verifiable AI Research: From Theorem Proving to Discovering New Structures

    AI research systems are moving beyond plausible-looking answers toward outputs that can be executed, checked, reproduced, and independently verified. This article examines a unified evaluation approach covering theorem proving, program execution, mathematical construction, optimization bounds, retry strategies, and research reporting. It also discusses current performance, system limitations, verification infrastructure, and the continuing role of human researchers.

    In recent years, artificial intelligence has made rapid progress in code generation, mathematical reasoning, and scientific computing. However, if evaluation remains focused on whether an answer merely resembles a correct one, it is difficult to determine how much transferable research capability a system actually possesses.

    Research-grade outputs must be reviewable, reproducible, and verifiable. Proofs should pass formal checking, programs should pass tests or performance evaluations, mathematical constructions should satisfy every constraint, and experimental conclusions should include clear records of inputs, procedures, and results.

    Against this background, a collaborative study involving AI researchers and mathematical theorists proposed a unified evaluation framework for research workflows. Rather than testing only whether a model can solve an individual problem, the approach combines problem understanding, solution generation, program execution, result verification, failure recovery, and final report generation within one inspectable system. This makes it possible to evaluate what an AI system can actually produce in realistic research tasks.

    Why Research Evaluation Needs to Be Verifiable

    Traditional question-and-answer benchmarks usually focus on whether the final text matches a reference answer. This is useful for factual questions, but it is not sufficient for mathematical proofs or algorithmic research.

    First, mathematical problems may have multiple equivalent answers. A model may reach the correct conclusion through a different approach without reproducing the standard solution. Second, local errors in a complex proof cannot always be detected from language alone. A seemingly reasonable lemma may fail under a boundary condition. Third, algorithmic research is often concerned with constructing an object that satisfies a set of constraints or improving an existing upper or lower bound, rather than producing a well-formatted explanation.

    A more reliable evaluation process generally includes the following stages:

    1. Task formalization: Define the inputs, constraints, objective function, and permitted proof tools.
    2. Candidate generation: Ask the system to propose a proof, program, parameter set, or structural construction.
    3. Automated verification: Use a formal proof checker, program tester, constraint solver, or independent computational implementation for validation.
    4. Result archival: Store the complete output, execution environment, random seeds, verification logs, and failure records.
    5. Human review: Have researchers assess the theoretical significance of the result and determine whether it genuinely improves on existing conclusions.

    The key value of this workflow is that it converts a sense of linguistic plausibility into research evidence that can be executed, inspected, and reproduced.

    What Capabilities Does a Unified Evaluation Cover?

    This type of research does not measure a single score. Instead, it attempts to cover several capabilities involved in scientific and mathematical work.

    1. Theorem-Proving Ability

    The system is asked to complete a set of theorems without being given the intermediate lemmas. Evaluation focuses not only on whether the final proof is complete, but also on whether it passes independent checking and avoids hidden assumptions.

    In one set of complete-proof tasks, the best-performing system solved 14 of 20 theorems, achieving a 70% success rate. This suggests that, when problem boundaries are clear, verification tools are available, and sufficient computation is allocated, AI can perform a substantial portion of formal mathematical work.

    However, the success rate should not be interpreted as a measure of general mathematical ability. The theorems were selected in advance, the input format was relatively structured, and the system could repeatedly search toward a clearly defined target. A more precise interpretation is that AI has demonstrated strong local execution capability in constrained proof-search tasks.

    2. Task Coverage Across Multiple Attempts

    Another evaluation allows the system to make multiple independent attempts and measures both the number of successful attempts and the number of distinct problems covered. One system succeeded 23 times in 60 attempts and covered nine different theorems, producing an overall success rate of 38.3%.

    This metric highlights an important point: research value cannot be measured by single-attempt accuracy alone. In open-ended research, it also matters whether a system can explore different approaches, identify failure patterns, and adjust its strategy in later attempts. A system with a modest per-attempt success rate but broader problem coverage may be more useful as a research assistant than one that performs consistently only on a narrow set of familiar problem types.

    3. Structural Construction and Combinatorial Optimization

    Another form of mathematical contribution is the direct construction of objects that satisfy complex constraints. For example, in an extremal graph task, a system constructed a graph with 124 vertices and 630 edges that contained neither triangles nor four-cycles. This exceeded the task’s previous reference value of 629 edges.

    The verification process for such a result differs from ordinary proof checking. The system must output the complete vertex set and edge set, while a verification program checks each of the following:

    • Whether the number of vertices is correct;
    • Whether the number of edges reaches the claimed value;
    • Whether duplicate or invalid edges are present;
    • Whether any three vertices form a triangle;
    • Whether any four vertices form a four-cycle;
    • Whether another program can independently reconstruct the output.

    The construction is credible only after every constraint has been satisfied. This shows that AI does not necessarily need to write a complete paper before contributing to research. A verifiable new construction can itself become the starting point for later theoretical analysis.

    4. Improving Upper and Lower Bounds

    In a coding-theory task, a system also proved that a particular parameter has an upper bound of at most 82, tightening the previously used public value of 84 to 82.

    Improving an upper bound generally requires stronger restrictions, more effective inequalities, or a new combination of known results. It is not simply a matter of computation. The system must identify new connections within an existing theoretical framework. Even an improvement of two units can matter because a tighter bound narrows the range of possible values and may help guide future work toward an exact solution.

    However, finding a smaller numerical value does not automatically constitute a mathematical result. Researchers must still inspect the proof chain, cited conditions, and scope of applicability to ensure that the improvement was not caused by a software bug, a missed boundary case, or inconsistent definitions.

    Results Show Uneven Capabilities Rather Than Universal Leadership

    In the evaluation, the two highest-ranked systems received overall scores of 71.9 and 57.4, while the highest score among the remaining systems was no more than 11.7. This indicates a substantial gap between systems, but it does not mean that one system dominates every task.

    For example, in a task requiring verification of the reliability of a coding alphabet, the second-ranked system completed all three attempts, while the top-ranked system failed to solve the task. Such local reversals are informative because they show that research ability is strongly task-dependent:

    • Some systems are better at long proof chains;
    • Some are more effective at searching discrete structures;
    • Some are more stable in program debugging and constraint satisfaction;
    • Some can suggest new directions but struggle with strict verification;
    • Some generate outputs quickly but are prone to boundary-condition errors.

    Overall scores therefore provide only a general reference. They cannot replace a detailed analysis of task types, failure causes, and verification costs.

    How These Systems Typically Work

    A relatively complete AI research-agent workflow can be abstracted as the following loop:

    1. Formalize the problem
    2. ↓
    3. Generate a proof, program, or construction
    4. ↓
    5. Call an independent verifier
    6. ↓
    7. Read error messages and counterexamples
    8. ↓
    9. Revise the approach and try again
    10. ↓
    11. Save reproducible experiment records
    12. ↓
    13. Have researchers review value and novelty

    The independent verifier is the core of the system. It may be a formal proof checker, a custom mathematical program, a constraint solver, a graph-structure detector, or independent code used to reproduce an experiment. The verifier does not need to understand the model’s intent. It answers a stricter question: does the output satisfy the conditions defined in advance?

    Several engineering practices are particularly important for improving reliability.

    1. Separate Generation from Verification

    A model should not generate an answer, judge its own answer, and then treat that judgment as final. The verification module should be as independent as possible to reduce the risk of repeating the same reasoning error.

    2. Preserve Failure Samples

    Research exploration naturally includes many failures. If only successful cases are saved, it becomes impossible to determine where the system actually breaks down. Complete logs should include the input, candidate answer, error messages, number of revisions, execution time, and resource consumption.

    3. Control Randomness and Resource Budgets

    Different attempts may be affected by random sampling, search depth, and execution time. Evaluations should specify the number of attempts, the time limit per attempt, memory limits, and the tools that may be called. Without these controls, results from different systems are difficult to compare.

    4. Prevent Verifier Exploitation

    If a verification program contains a vulnerability, a system may obtain a false success through special formatting, malformed inputs, or boundary cases that the verifier does not cover. Important results should therefore be checked with multiple independent implementations whenever possible, and all inputs should be normalized strictly.

    The Remaining Boundaries of AI Research

    Although these results are encouraging, they are still far from showing that AI can conduct research autonomously.

    First, humans usually select the tasks. A system may be good at solving a given problem without knowing which problems are worth pursuing or reliably determining whether an improvement will have lasting significance.

    Second, the value of a result requires scholarly judgment. A new construction may be an isolated special case, or it may reveal a deeper pattern. Assessing its importance requires knowledge of the field’s history, problem context, and unresolved challenges.

    Third, open problems rarely provide clear feedback. For tasks with a standard verifier, a system can iterate based on error messages. In genuinely frontier research, there may be no obvious failure signal and no existing program that can determine whether a direction is promising.

    Fourth, complex research requires collaboration across stages. Problem selection, literature review, conjecture generation, proof design, experimental validation, and presentation depend on one another over long periods. A single successful result does not show that a system can consistently complete an entire research chain.

    Implications for Developers and Research Teams

    For organizations seeking to introduce AI into scientific or engineering workflows, the highest priority should not be a more sophisticated chat interface. It should be reliable verification infrastructure.

    Practical steps include:

    • Decompose each problem into inputs, constraints, objectives, and acceptance criteria;
    • Design automated checkers for code, mathematical objects, and experimental results;
    • Preserve complete intermediate artifacts instead of storing only final answers;
    • Use repeated generation and verification rather than one-shot output;
    • Train error classification and strategy-selection capabilities with failure cases;
    • Report model scores separately from practical research value;
    • Independently reproduce and manually review important conclusions.

    From an engineering perspective, future research-oriented AI systems will look more like collaborative platforms composed of models, execution environments, verifiers, task queues, logging systems, and human review. They will not be simple text generators. The model will propose candidate solutions, tools will execute them, verifiers will make formal judgments, and researchers will choose directions and interpret significance.

    Conclusion

    The most important change demonstrated by this evaluation is not that AI can already replace mathematicians or program researchers. It is that AI is beginning to produce genuine results in verifiable portions of the research process. Whether completing some theorems, discovering new combinatorial structures, or tightening an existing theoretical bound, these examples show AI moving from content generation toward the generation of reviewable evidence.

    A more accurate conclusion is that AI can now execute parts of research tasks when the goal is clear, the computational budget is sufficient, and automated feedback is available. Choosing research directions, judging the value of results, and making breakthroughs on major open problems still depend on human experience, intuition, and long-term accumulation of knowledge.

    The most promising future path is therefore not simply to build larger models. It is to create a tighter loop connecting models, tools, verifiers, and researchers.