Assessing LLM Research Capabilities through Verifiable Benchmarks
Public benchmark scores do not necessarily reflect an LLM's ability to conduct original research. This article explores how to evaluate model research capabilities using a verifiable test set based on Theoretical Computer Science (TCS). It covers the limitations of standard benchmarks, the design principles for building verifiable evaluation datasets—including task categorization and anti-contamination measures—and provides guidance on interpreting results and building private evaluation sets for enterprise applications.
1. Why Public Leaderboard Scores Fall Short
The pace of new model releases is rapid, and public leaderboard scores are frequently updated. However, high scores do not mean a model can conduct research. Common benchmarks resemble closed-book exams: questions are drawn from known problems, solutions follow existing paradigms, and grading can be automated. This design measures a model’s ability to reproduce and combine existing knowledge, which is fine in itself. The problem arises when we use it to infer scientific research capabilities, often leading to overly optimistic conclusions.
True research activities have three distinct characteristics. First, the problem-solving path is unknown, and methods often need to be invented. Second, the outcome is unknown; sometimes, even the person asking the question is unsure if an answer exists. Third, the value is hard to judge—even if a conclusion is provided, determining whether it is correct and valuable requires significant time from experts.
Inferring research capability from exam scores is like predicting whether someone can invent a new culinary genre based on their cooking proficiency exam results. The two are related but not equivalent.
2. Grading is Harder than Generating Questions
There is a core contradiction in evaluation design: the closer a question is to real research, the harder it is to grade; the easier it is to grade, the less it resembles research. If a model claims to have proven a long-standing conjecture, the verifier must check the proof step by step. Manual review is costly, and the time of top experts is highly limited. If expert reviews must be organized for every new model release, the evaluation itself becomes unsustainable.
Therefore, when designing an evaluation set, one must first answer a question: can the correctness of this problem be checked in a deterministic way?
3. Why Theoretical Computer Science is a Suitable Carrier
Theoretical Computer Science (TCS) studies the boundaries of computation itself. Does a problem have an efficient algorithm? Where are the complexity lower bounds? Which problems are fundamentally uncomputable? Unlike engineering implementations, these questions rely more on rigorous mathematical proofs.
A subset of conclusions in TCS possesses machine-verifiability, which is the primary reason it is suitable as an evaluation carrier:
- Formalized proofs can be verified step-by-step by proof checkers, independent of the reviewer’s intuition.
- Results such as counterexamples, constructions, and reductions can be written as programs and tested.
- Correctness determination is deterministic, leaving no room for ambiguity.
It should be noted that not all TCS propositions can be graded by machines. Many cutting-edge papers rely on intuition, engineering experience, or unformulated arguments. Therefore, these conclusions apply only to the machine-verifiable subset and cannot be generalized to the entire discipline.
4. How to Build the Evaluation Dataset
A typical evaluation set can categorize questions based on the type of research activity. Four common categories are:
- Proofs: Given a proposition, the model is required to complete a rigorous proof or disprove it.
- Constructions: The model must design an algorithm, protocol, or structure satisfying specific properties.
- Analysis: The model must derive the complexity, correctness, and boundary conditions of a given scheme.
- Open-ended Exploration: Problem variants close to the real research frontier, testing the model’s exploratory boundaries.
The grading methods for these four categories differ. Proofs can be submitted to proof checkers or manually reviewed. Constructions can be checked by writing test programs to verify if the output meets the properties. Analysis can require the model to provide verifiable intermediate conclusions, such as algorithm examples at specific scales. Open-ended exploration is difficult to grade fully automatically; in practice, it is usually broken down into verifiable sub-goals, or formalized variants with known answers are constructed, with the remaining parts left to expert review.
Three principles should be followed during design:
- Verification First: Each question must specify its grading method. Questions that cannot be automatically graded should be marked as “expert review” and not mixed into automated scoring.
- Difficulty Stratification: Form a capability profile ranging from graduate course problems to open problem variants, avoiding a single aggregate score.
- Anti-Data Contamination: Questions must be originally modified. Use time truncation to check the cutoff point of the model’s training data, and use deduplication and n-gram overlap detection to see if the original text of the question has appeared in public corpora. Questions with high overlap need to be replaced or manually reviewed.
5. How to Interpret the Results
Highly structured questions with mature solution paradigms usually achieve higher pass rates. When entering areas requiring original insight and long chains of rigorous reasoning, the success rate drops significantly. On the most difficult open problems, under the evaluation setting referenced in this article, no participating model was observed to solve the problem independently. This conclusion applies only to this specific setting and cannot be directly extrapolated to all models and all problems.
When interpreting results, note the following:
- Specify the participating models, number of questions, grading criteria, and definition of failure; otherwise, scores are incomparable.
- Distinguish between “providing an answer” and “the answer passing verification”; the gap between the two can be large.
- Focus on the stratified profile rather than just the total score.
A phenomenon worth alerting appeared during the evaluation: models would generate proofs that appear rigorous, well-symbolized, and logically fluent, but upon closer inspection, contain flaws in key steps. Such proofs are harder to deal with than simply admitting a lack of knowledge, as they increase the cost of manual review. During review, the following checklist can be checked item by item:
| Check Item | Specific Question |
|---|---|
| Key Lemmas | Is every cited conclusion well-founded and actually true? |
| Boundary Conditions | Are empty inputs, minimum scales, and degenerate cases handled? |
| Quantifier Order | Is the order of “for all” and “exists” incorrectly swapped? |
| Counterexample Construction | Can a small example be found to test if the argument fails? |
6. Recommendations for Developers and Enterprise Model Selection
This methodology has direct reference value for engineering practices. The following table provides recommendations based on use cases:
| Scenario | Verifiable Means | Review Requirements | Risk Warning |
|---|---|---|---|
| Coding, documentation, routine algorithm implementation | Unit tests, static analysis, type checking | Routine code review | Passing tests does not guarantee correctness if test coverage is insufficient |
| Scheme argumentation, correctness proof, boundary deduction | Formal verification, counterexample testing, property testing | Expert review is mandatory | Fluent arguments may hide flaws |
| Open-ended technical exploration | Break down into verifiable sub-goals | Primarily manual judgment | Model conclusions may not be independently verifiable and are not suitable for direct deployment |
Before introducing model capabilities, a small private evaluation can be built in three steps:
- Task Sampling: Extract representative tasks from real business operations, covering common and high-risk scenarios.
- Answer Verifiability: Design a grading method for each task. Those that can be automatically graded should be written as tests, while those that cannot should be marked for manual review.
- Regression Monitoring: Rerun the evaluation after every model change or prompt modification, recording score changes and comparing them with the previous version.
Private evaluation sets contain business data, so data isolation must be maintained, access restricted, and sensitive samples should not be sent to uncontrollable third-party services. Running evaluations also incurs costs, including inference fees, manual review time, and maintenance expenses. These overheads should be estimated before scaling up.
7. Conclusion
The primary value of this evaluation approach is providing a verifiable measurement methodology. It separates “solving known problems” from “exploring the unknown,” and ensures every conclusion has a clear scope of application.
For mathematicians and theoretical scientists, this indicates that current models still have significant limitations in original research. For engineers, the practical approach is: use models confidently in verifiable steps, remain cautious in steps requiring insight, and build evaluations using your own tasks. The core question the evaluation must answer is simple: where is the model best suited?
