Evaluating LLM Research Capabilities: How to Build Trustworthy Benchmarks

    Leaderboard scores mostly measure how well a model takes standardized tests, not how well it conducts open-ended research. This article examines why "solving problems" and "doing research" are fundamentally different, why credible grading is the hardest part of research-level evaluation, and how machine-verifiable tasks—especially in Theoretical Computer Science (TCS)—offer a practical foundation for trustworthy benchmarks. It outlines the design principles for such benchmarks and gives enterprises and developers concrete guidance on evaluating and selecting models.

    1. Behind the Crowded Leaderboards: What They May Only Prove Is “Test-Taking”

    Over the past two years, LLM leaderboards have been refreshed almost every month or two. From general knowledge Q&A to math word problems, from code generation to logical reasoning, benchmark after benchmark has been released and scores have kept climbing. These leaderboards give the industry an intuitive reference, but they also foster a common misconception: that the higher the score, the closer a model is to a researcher’s level.

    The problem is that the vast majority of mainstream benchmarks are essentially “standardized exams.” Problems come with clear given conditions, relatively fixed solution steps, and a finite set of standard answers. Scoring high on such tests says more about a model mastering a particular problem-solving paradigm than about having the ability to independently explore the unknown. It’s like cooking from a recipe: the steps are ready-made, the heat is labeled, and the flavor has a reference standard. Someone who can cook may still be unable to invent a new dish.

    2. “Solving Problems” vs. “Doing Research”: A Gap Hard to Cross Automatically

    Real research differs fundamentally from problem-solving:

    • No ready-made recipe: Research problems often have no prescribed solution path, and in many cases even the person posing the problem doesn’t know whether an answer exists.
    • Uncertain process: Whether a problem can be solved, and how many detours it takes, cannot be predicted in advance.
    • Open-ended results: The output of research is often a new conjecture, a new proof, or a new counterexample, rather than recomputing a known answer.

    As a result, a model that scores near-perfect on competition problems may not be able to take even a small step on an open problem. To measure a model’s research ability, we can’t just look at whether it “can solve problems”; we have to examine how it actually performs on open-ended problems with no standard answer.

    3. The Grading Problem: Why “I Solved It” Isn’t Trustworthy

    When evaluating research ability, the trickiest part is grading.

    When a model claims to have cracked an open conjecture that has stood for decades, or produces a lengthy proof, who verifies whether it’s correct? This is precisely the risk most easily overlooked in the “AI does math” narrative: generating a rigorous-looking piece of reasoning text is easy, but making that text stand up to strict scrutiny is hard. A model may earnestly write out logical flaws, circular arguments, or fabricated conclusions, and such errors are often difficult to verify line by line.

    Without a trustworthy judging mechanism, so-called “breakthroughs on century-old problems” may remain at the level of self-declaration. The credibility of evaluation depends on whether the verification process itself is reproducible, automatable, and adversarial.

    4. The Path Forward: Betting Evaluation on “Machine-Verifiable”

    A natural solution is to only pose problems whose conclusions can be automatically judged right or wrong by a machine.

    The closer a problem is to real research, the harder it usually is to score automatically; but in some disciplines, there exists a class of problems that are “hard to generate, easy to verify”—producing an answer may require extraordinary creativity, while checking whether the answer is correct can be done efficiently by a program. This asymmetry is the fulcrum for building trustworthy research-grade evaluation. It’s a bit like Sudoku: filling in a correct final grid may be taxing, but once it’s filled, checking whether any rules are violated takes a second.

    Relying on such verifiable tasks, evaluation can break free from the bottleneck of “depending on expert manual review” and become scalable, standardized, and reproducible, while still retaining the strength to assess genuine research ability.

    5. Theoretical Computer Science: A Natural Verifiable Testbed

    Among all disciplines, Theoretical Computer Science (TCS) is especially suited to this task.

    TCS doesn’t study how to fix machines or tune code; it studies the boundaries of “computation” itself: how fast can a problem be solved at best? Are there classes of problems that are inherently impossible to solve efficiently? The classification of complexity classes, the limits of approximation algorithms, the proofs of lower bounds—these results are mostly presented in the form of mathematical proofs.

    And proofs have a key property: their correctness can be formally checked. As long as the reasoning process is expressed as rigorous formal steps, it can be verified step by step with the help of automatic or semi-automatic theorem-proving systems, with almost no reliance on the subjective judgment of the evaluator. This makes TCS both open-ended in its research problems and verifiable in its answers—exactly the ideal evaluation material.

    6. How a Research-Grade Benchmark Should Be Designed

    Following this idea, a research-ability evaluation for LLMs should typically cover several typical tasks in the daily work of theoretical computer science researchers, for example:

    • Given a problem, construct a correct solution or proof;
    • Propose new, valuable conjectures building on existing results;
    • Find counterexamples or constructive instances for a proposition;
    • Identify and correct errors in proofs.

    In design, a few principles need to be kept in mind:

    1. Verifiability first: Prioritize problems whose conclusions can be judged by a machine, ensuring objective and reproducible scoring.
    2. Close to real research: Problems should faithfully preserve openness and difficulty rather than degenerating into formulaic exercises.
    3. Assess by dimension: Evaluating “problem-solving,” “conjecturing,” “construction,” and “error-correction” separately is the only way to see a model’s capability structure clearly.
    4. Resistant to data contamination: By controlling the source and timeliness of problems, try to prevent models from “memorizing answers” from training data.

    It’s worth emphasizing that the goal of evaluation is not to give a model an isolated score, but to characterize its capability boundaries: which tasks it can already handle, and where it still makes errors it can’t self-correct. Viewing these conclusions rationally is more valuable than chasing a pretty ranking.

    7. Implications for Enterprises and Developers

    For enterprises and technical teams, this kind of “verifiable, multi-dimensional” evaluation method offers a more pragmatic perspective for model selection:

    • Don’t just look at the overall leaderboard score; look at a model’s stability and reliability on specific task dimensions.
    • For high-value, high-risk scenarios (such as analysis and reasoning that require strict logical correctness), prioritize tasks that can be automatically verified for pre-launch validation.
    • With general-purpose cloud model services, compute resources, and monitoring and alerting capabilities (for example, the model services and engineering support provided by Baidu AI Cloud), turn evaluation, verification, and iteration into a sustainable engineering process rather than a one-off score.

    Back to the original question: how strong is the research ability of LLMs? A more prudent answer is—it’s progressing rapidly at “solving problems,” but when facing truly unresolved open problems, we still need to stay clear-headed. Building verifiable research-grade benchmarks is precisely about grounding this judgment on solid evidence, rather than the noise of trending topics and leaderboards.