What a Theoretical Computer Science Exam Reveals About the Limits of LLM Research Capabilities

    A new benchmark called TCSExam evaluates large language models on tasks that resemble real theoretical computer science research, including theorem understanding, algorithm design, formal proof construction, and open-ended conjecture exploration. The results draw a clear line between reproducing established knowledge and making genuinely original research progress: while leading models perform well on structured tasks, none independently solved any of the open problems in the evaluation. The benchmark also shows why machine-verifiable tasks may be essential for measuring AI research ability objectively.

    Introduction: How Far Is AI from Revolutionizing Mathematics?

    Debate over whether AI is transforming mathematical research has intensified over the past year. Several research organizations have publicly released hundreds of AI-assisted mathematical manuscripts covering number theory, theoretical computer science, geometry, mathematical physics, and more. The topic has attracted considerable public attention.

    At the same time, prominent mathematicians have pointed to a fundamental weakness of AI in almost every knowledge-intensive field: models can produce errors that are difficult to detect. Mathematics appears to be a notable exception because the correctness of a proof can, at least in principle, be checked rigorously by machines. This is one reason many researchers expect AI’s least controversial achievements to emerge first in mathematics.

    A recently published study, however, offers a more cautious perspective. The study was conducted by several researchers in theoretical computer science, including a Turing Award winner and the founder of a large-model company, along with researchers from academic institutes. Together, they created an evaluation called TCSExam, or the Theoretical Computer Science Research Evaluation, to test several leading large language models.

    The basic idea is simple: the researchers created an exam designed to resemble real research work and gave it to AI systems. The results were far from encouraging. None of the evaluated models independently solved any of the still-open conjectures included in the exam.

    Why Theoretical Computer Science?

    An Analogy: Drivers and Transportation Planners

    Theoretical computer science, or TCS, is not primarily about writing better code. It studies the fundamental laws of computation itself.

    A useful analogy is the difference between a driver and a transportation planner. A programmer is like a driver who cares about engine performance, the driving experience, and the conditions on a particular road. A theoretical computer scientist is more like a transportation planner who asks how quickly it is possible to travel from one place to another, whether some routes are fundamentally impossible, and whether a general solution exists for an entire class of problems.

    TCS lies at the intersection of mathematics and computer science. Its questions revolve around computation, and its main research method is rigorous proof. Core areas include algorithm design and analysis, lower bounds in computational complexity, graph theory, logic and automata theory, and the theoretical foundations of cryptography and game theory.

    Machine Verifiability Makes TCS a Natural AI Testbed

    There is a practical technical reason to evaluate AI systems on TCS: many of its conclusions can be verified automatically by machines.

    Checking a mathematical proof is somewhat like checking a completed Sudoku puzzle. Filling in the solution may be difficult, but once it is complete, the result can be checked against a set of rules. Similarly, whether an algorithm satisfies its correctness requirements, whether a complexity claim holds, and whether every intermediate step in a constructive proof can be formalized are all tasks that can potentially be verified step by step.

    This makes TCS particularly suitable for automated grading. By contrast, writing, question answering, and code generation can also be scored, but their evaluations are generally less objective and less stable than tasks involving formally verifiable proofs.

    That distinction is critical because grading is one of the central challenges in AI evaluation.

    Why Existing Benchmarks Are Not Enough

    Solving Problems Is Not the Same as Doing Research

    Large language model benchmark scores are updated constantly, and the growing number of leaderboards can be difficult to interpret. Yet an increasingly common conclusion is that these benchmarks mainly show whether a model can solve problems, not whether it can conduct research.

    The difference can be illustrated as follows:

    • Problem solving is like cooking from a recipe. The steps already exist, the result has a defined standard, and there is little room for an unexpected but valid answer. Because training data often contains similar problem types and fixed grading criteria, a model may arrive at the correct answer without fully understanding the underlying principle.
    • Research has no prescribed sequence of steps. No one knows in advance whether an approach will succeed. For many open problems, even the person posing the question does not know the answer, so there is no standard solution for a model to imitate.

    As a result, a model that achieves a high score on a difficult problem set may still be unable to generate useful ideas for an open problem without prompts, references, or a known answer.

    Existing benchmarks also face three major issues: possible data contamination, repetitive task formats, and subjective grading. For these reasons, an increase in benchmark scores cannot automatically be treated as evidence of improved research capability.

    The Grading Problem: When an AI Claims to Have Solved a Conjecture

    The challenge becomes even more serious when a model claims to have solved a conjecture that has remained open for decades. Who can verify that the proposed proof is actually correct?

    Publicly released AI-generated mathematical manuscripts provide a clear example. They inevitably contain many arguments that appear plausible but are incorrect. Checking each one in detail is extremely time-consuming. Unless the feasibility of grading is addressed, AI-generated “research results” cannot be evaluated with the same level of seriousness as human research.

    One possible strategy is to focus on problems that machines can grade. The closer an evaluation gets to real research, however, the harder it becomes to score. TCS is a particularly useful exception because many of its conclusions can be machine-verified. This forms the methodological foundation of TCSExam.

    Exam Design: Four Tasks That Reflect Research Practice

    The evaluation contains 100 questions divided into four categories based on the typical activities of theoretical computer scientists.

    1. Understanding Established Results

    These tasks ask models to explain, restate, and derive conclusions from theorems already published in the literature.

    They test how deeply a model understands domain knowledge and whether it can reconstruct a complex chain of reasoning within a limited context. The requirements for context handling and logical consistency are demanding: one incorrect inference can cause the entire subsequent argument to fail.

    2. Algorithm Design and Analysis

    Given a specific problem, the model must design a correct algorithm and provide a rigorous time-complexity analysis.

    This category is closest to the everyday work of many engineers, but its difficulty lies in the required level of rigor. It is not enough for an algorithm to appear to work. The model must prove that it is correct for every valid input. Arguments showing that a complexity bound cannot be improved also require substantial theoretical knowledge.

    3. Proof Construction

    In this category, models receive a proposition and must produce a complete formal proof.

    This is the most difficult part of the exam because it directly simulates the central activity of research: turning intuition into a logically complete argument without gaps.

    The models tended to fail in similar ways. They introduced assumptions that had not been verified, overlooked boundary cases in constructions, or relied on circular reasoning that initially appeared plausible.

    4. Exploring Open Problems

    For unresolved conjectures, models were asked to propose well-founded ideas or make partial progress.

    The outcome matched the broader conclusion of the study: none of the evaluated models independently solved any of the open conjectures. Most of their proposed ideas were high-level restatements of existing work rather than genuinely original advances.

    Key Conclusions and Implications

    Correcting the Claim That AI Is “Defeating Mathematics”

    The evaluation reveals a clear boundary for current frontier models. They can competently reproduce and apply established theoretical results, and on tasks with recognizable patterns they may perform at a level approaching or exceeding that of human graduate students.

    However, they remain far from independently completing research tasks that require original exploration, rigorous construction, and progress on open questions.

    Claims that AI has already “destroyed mathematics” are therefore better understood as attention-driven exaggerations than as accurate descriptions of current capabilities.

    Three Lessons for the Industry

    First, research evaluations need to evolve. Research-oriented benchmarks should be built around verifiability. Tasks whose conclusions can be checked by machines are more likely to provide objective measurements of AI research ability.

    This has direct implications for companies developing model evaluation systems. Rather than relying primarily on subjective leaderboards, organizations can design task sets whose outputs are subject to formal validation.

    Second, verifiability is a prerequisite for scalable deployment. Mathematics and theoretical computer science are among the fields where AI capabilities can be demonstrated most clearly because their results are naturally testable.

    In enterprise applications, constraining AI outputs to domains where they can be automatically checked—such as code generation evaluated with test cases or reasoning tasks governed by formal validation rules—can help control hallucination risks and produce more reliable results.

    Third, human-AI collaboration remains the most practical path. The evaluation does not mean that AI has no value in scientific research. As an assistant for literature review, idea generation, and early-stage drafting, large language models can already reduce researchers’ preparation costs significantly.

    A more realistic near-term direction is not to expect AI to complete an entire proof independently. Instead, models should first become reliable at producing small, verifiable steps that can be checked automatically, while human researchers remain responsible for the critical creative leaps.

    Conclusion

    A 100-question exam does not merely test how much knowledge a model has memorized. It tests whether the model can reason like a researcher.

    For now, the answer is no. Yet the evaluation itself is significant because it provides one of the first relatively credible measurements of AI research capability based on machine-verifiable tasks.

    As competition among large language models becomes increasingly intense, careful, rigorous, and reproducible evaluations such as TCSExam may be more valuable to the industry than another record-breaking benchmark score.