When LLMs Start Doing Research: Lessons from a Machine-Gradeable, Full-Domain Evaluation

    A new benchmark in theoretical computer science—led by a Turing Award laureate—uses machine-verifiable problems to test whether frontier LLMs can actually do research. The result: no model could independently solve any open conjecture. This article breaks down the evaluation's design, why machine-gradeable scoring matters, and what engineers can learn when building their own model assessments.

    When LLMs Start Doing Research: Lessons from a Machine-Gradeable, Full-Domain Evaluation

    Introduction: Do High-Benchmark LLMs Really Know How to Do Research?

    In recent years, large language models have been breaking benchmark records every month or two. Leaderboards keep shifting, creating the impression that model intelligence is visibly approaching—or even surpassing—human experts. Yet one fundamental question remains unresolved: what do these high scores actually prove?

    A recent piece of academic work carries real weight here: a team of scholars from universities and frontier research institutions, led by a Turing Award laureate, released a full-domain research-capability evaluation for theoretical computer science (TCS). They built a 100-question exam covering the field’s four major research directions and used it to test the strongest available models. The conclusion poured cold water on the hype—when facing open conjectures that remain unsolved to this day, not a single participating model could solve them on its own.

    This work deserves the attention of technology practitioners not only because of the authors’ academic standing, but because it introduces a genuinely new evaluation paradigm: using machine-gradeable questions to assess “doing research”—an ability that is inherently resistant to standardized scoring. This article unpacks the logic, methodology, and engineering lessons behind that idea.

    1. Solving Problems vs. Doing Research: Two Fundamentally Different Abilities

    1.1 Problem-Solving Is “Cooking from a Recipe”

    Most benchmarks measure problem-solving ability. These problems share three characteristics:

    • Known path: No matter how hard the problem is, the solution steps already exist—someone has paved the way;
    • Standard answers: A grader simply compares the output against a reference answer;
    • Closed boundaries: The problem is well-defined, with no dispute over whether it is even valid.

    This is like cooking a dish strictly from a recipe: oil temperature, batter thickness, seasoning ratios—everything is prescribed. The model only needs to retrieve, recombine, and execute. Even when it performs beautifully, it remains an advanced form of pattern matching.

    1.2 Research Is “Cooking Without a Recipe”

    Real research is something else entirely:

    • No prescribed steps—researchers must decide for themselves which path to take;
    • Unpredictable outcomes—the problem itself may even be ill-posed;
    • For many open problems, nobody knows the answer—and that is precisely the norm in research.

    If a model claims to have cracked a conjecture that has stood for decades, who verifies that its proof is correct? Verifying a cutting-edge proof often takes top scholars in the same field weeks or even months. An AI company once publicly released hundreds of AI-generated mathematical manuscripts at once, sparking widespread debate about whether AI had already replaced mathematicians—but the validity and depth of those results lacked any systematic means of verification.

    1.3 Why Math and Theoretical Computer Science Became the Testing Ground

    Among all disciplines, mathematics and TCS have a unique property: the correctness of a proof can be checked automatically by machines. It is a bit like Sudoku—hard to fill in, but trivial for a machine to verify once completed. Formal verification tools (such as interactive theorem provers) can decompose every step of a proof into a logical chain checkable by mechanical rules.

    That is why AI results in mathematics are the easiest to make “indisputable.” It also explains why frontier research institutions have converged on math and theoretical computer science as the touchstone for testing a model’s research ability.

    2. Why a New Exam Was Needed

    2.1 Three Ways Existing Benchmarks Fail

    Traditional evaluation systems fail systematically when it comes to measuring research ability:

    1. Saturation failure: Many classic benchmarks have been pushed close to perfect scores, losing discriminative power—it is impossible to tell whether a model has hit the ceiling or just a paper-thin wall;
    2. Contamination failure: Many questions already appear in public training corpora, so high scores may come from memorization rather than reasoning;
    3. Category failure: Benchmark performance and research ability are simply not the same thing; inferring the latter from the former is a category error.

    2.2 The Core Design Principle: Machine-Gradeable Questions

    The most elegant design decision of this TCS evaluation is: only include questions that machines can grade.

    There is a natural tension here—the more a question resembles real research, the harder it is to grade automatically; and the harder it is to grade automatically, the more the conclusions can be challenged. TCS happens to be an exception: most of its conclusions are machine-verifiable. By carefully selecting question types, the authors found a balance between “research authenticity” and “scoring objectivity.”

    The lesson for engineering is universal: any AI capability evaluation whose scoring depends on subjective human judgment cannot scale and cannot be reproduced. When designing an evaluation system, “automatically verifiable” should be treated as a first-class requirement.

    2.3 Four Question Types, Mirroring a Researcher’s Four Daily Activities

    The exam contains 100 questions, organized into four categories that map to how theoretical computer scientists actually work:

    • Proposing and proving new theoretical results: testing the creative ability to build proofs from scratch;
    • Analyzing the properties of a given proposition: testing the ability to dissect and reason about complex theoretical structures;
    • Solving open or semi-open problems: testing the ability to explore uncharted territory;
    • Verifying and constructing computational instances: testing the ability to support theoretical judgments through concrete constructions.

    This “categorize by professional activity” approach is far superior to categorizing by knowledge points—it evaluates the structure of capability rather than knowledge coverage. Engineers designing evaluations can borrow the same idea: first model the real workflow of the target role, then design a verifiable task for each stage.

    3. Where the Cold Water Lands: Collective Failure on Open Problems

    The most notable finding of the evaluation is this: facing open conjectures that remain unsolved, no participating model could independently produce a proof.

    This result is not surprising, but it is highly valuable. It draws a clear boundary around what current LLMs can actually do:

    • Long-horizon reasoning along known paths: models are getting steadily stronger, capable of multi-step, long-chain deduction;
    • Exploratory breakthroughs along unknown paths: models have no reliable ability here; their “creativity” mostly shows up as recombination of existing knowledge, not genuine paradigm shifts.

    Mechanistically, this makes sense: an LLM is fundamentally distribution learning over massive corpora. It excels at operating near the “high-probability ground that humans have already walked.” Solutions to open problems, by contrast, tend to lie in the sparse regions of that distribution—requiring counterintuitive leaps, questioning of established frameworks, and long periods of exploration without a clear target. None of these fall within the reach of the current training paradigm.

    That said, this does not mean models are useless for research. A more accurate framing is: today’s LLMs are “advanced assistants” to researchers, not replacements—they can accelerate literature surveys, assist formal verification, generate candidate ideas, and batch-check lemmas, but the step of producing the key insight still belongs to humans.

    4. Engineering Takeaways for Technology Practitioners

    For developers, this evaluation offers not just academic news but actionable methodology:

    4.1 Build Your Own Verifiable Evaluation Set

    If you are responsible for model selection or performance evaluation in a company, you can adopt the same approach:

    1. Evaluation set design checklist:
    2. 1. Does every question have an objectively machine-gradeable answer?
    3. 2. Has each question been confirmed absent from public training corpora?
    4. 3. Do the questions cover the real task types of the target scenario, not just knowledge points?
    5. 4. Is the scoring script reproducible and versionable?
    6. 5. Does the set include "probe" questions beyond current capabilities, to observe progress over time?

    4.2 Beware of “Leaderboard Illusions”

    High scores on third-party leaderboards do not equal high usability in your business scenario. The right approach: use public leaderboards for initial screening, use a self-built, auto-graded business evaluation set for final selection, and continuously update the question bank with data flowing back from production.

    4.3 Automate Verification First

    In applied scenarios such as AI-assisted coding and knowledge Q&A, any step whose output can be programmatically checked—unit tests, type checking, constraint solving—should be wired into a verification pipeline first. The model “generates candidates”; the program “adjudicates correctness.” This division of labor follows the same design philosophy as the TCS evaluation and is the most reliable engineering paradigm at the current stage.

    4.4 Use General-Purpose Cloud Infrastructure to Support the Evaluation Pipeline

    Building large-scale evaluations requires elastic compute scheduling, unified logging, and monitoring/alerting capabilities. Building the evaluation pipeline on general-purpose container platforms and object storage makes question-bank versions, inference results, and scoring reports traceable end to end—providing a solid data foundation for model iteration.

    Conclusion: What Reassures Researchers Is Not That AI Falls Short, but That the Boundary Is Clear

    The reason this evaluation let mathematicians and theoretical computer scientists “breathe a sigh of relief” is not schadenfreude—it is that the boundary has been drawn clearly. Knowing what AI can and cannot do is more valuable than vague panic or blind optimism. For the engineering community, this is an equally important reminder: while chasing leaderboard numbers, we must also build our own capability yardsticks—verifiable ones, grounded in real tasks. Only then can we put the tool to work in the right place in this new era of human-machine collaboration.