Evaluating AI Research Agents: Capability Boundaries of Verifiable Closed Loops
This article explores evaluation frameworks for AI research agents, shifting the focus from generating answers to producing verifiable artifacts. It examines the transition toward generating certificates, the spectrum of research tasks, and the design of metrics that better reflect real-world research. Additionally, it discusses the boundaries of application, the division of labor between humans and AI, and how enterprises can leverage verifiable closed loops in R&D pipelines using general cloud capabilities.
Over the past few years, the vision of large language models evolving from “answering questions” to “conducting research” has been continuously magnified. However, the real challenge is not generating a seemingly profound piece of text, but producing an object that can be independently verified by a third party: a proof, a construction, a better program, or a set of reproducible experimental parameters. Recently, a class of evaluation frameworks oriented toward research agents has garnered attention. Their value lies not in the ranking of a specific system, but in breaking down the research process into verifiable segments: problem formalization, candidate solution generation, certificate output, independent verification, budget recording, and failure attribution. In other words, they attempt to answer: given clear goals, sufficient compute, and verifiable feedback, which parts of the research process can AI actually undertake, and where must it still rely on human judgment?
From “Generating Answers” to “Generating Certificates”
Traditional QA evaluations focus on whether the final answer matches, while research scenarios emphasize whether the “evidence chain is closed.” A useful conclusion should be accompanied by a verifiable artifact: mathematical proofs must be verifiable by proof checkers or step-by-step human review; graph constructions must provide vertex sets, edge sets, and forbidden subgraph judgments; coding bounds must provide check matrices or search logs; program optimizations must pass unit tests, property tests, and differential testing. The significance of a certificate lies in transferring trust from the model’s identity to the verification process; even if the generator makes a mistake, the verifier can still reject the erroneous conclusion.
The schematic process can be abstracted as follows:
task = formalize(problem_statement)budget = allocate(search_time, memory, tool_calls)for attempt in range(budget.max_attempts):candidate = agent.propose(task, context)certificate = agent.emit_witness(candidate)verdict = verifier.check(task, candidate, certificate)ledger.record(attempt, verdict, cost, artifacts)if verdict == VALID:breakreport = aggregate(pass_at_k, coverage, novelty, cost)
The key here is not how powerful the model is, but that the verifier must be as independent as possible from the generator. The same set of propositions cannot be evaluated solely by the generation model; otherwise, “linguistic fluency” would be misjudged as “logical validity.”
Task Spectrum: Proofs, Constructions, Optimization, and Discovery
These evaluations typically cover four categories of tasks. The first is theorem proving, where the goal is to complete formal or semi-formal derivations given a set of axioms and a lemma library. The difficulty lies in long-chain dependencies, intermediate lemma selection, and search space explosion. The second is combinatorial construction, such as finding more edges given girth constraints in extremal graph theory, or tightening parameter upper bounds in coding theory. Its value can be characterized by “exceeding reference constructions” or “tightening published bounds,” but it must be recomputed by an independent program for edge counts, forbidden cycles, distance spectra, or verification conditions. The third is program and system optimization, requiring the reduction of latency, memory, or complexity while maintaining semantic equivalence, verified through benchmarks and regression tests. The fourth is open-ended exploration, including proposing conjectures, choosing directions, and judging value. This level is closest to real research and is currently the weakest part of automation.
Public experiments reveal similar patterns: on tasks with clear feedback and complete verifiers, leading systems can reliably solve a considerable portion of problems. On tasks lacking intermediate lemmas, requiring cross-domain transfer, or needing value judgments, the success rate drops rapidly. Some systems also exhibit “specialization” in specific subtasks—one party might complete all reliable constructions of a certain alphabet, while a generally stronger system fails. This indicates that a total score cannot replace a capability profile, and leaderboard rankings cannot be directly equated with a research replacement rate.
Metric Design: Closer to Research Reality Than pass@1
If only the success of a single attempt is counted, the model will be overestimated or underestimated. More reasonable metrics should include: pass@k, measuring reachability under multiple samplings; coverage, counting how many different theorem or construction families are covered rather than repeatedly hitting the same problem; novelty delta, comparing the improvement margin relative to reference constructions or published bounds; cost-normalized success, dividing the number of successes by search steps, tokens, tool calls, and wall-clock time; lemma dependency, recording whether intermediate lemmas, external libraries, or human prompts are allowed; and failure taxonomy, distinguishing between syntax errors, verification timeouts, unsatisfied construction constraints, proof gaps, and goal misunderstandings.
Engineering-wise, it is also necessary to prevent “evaluation overfitting”: problems cannot come from explicit answers in the training corpus; reference solutions need variant and isomorphic rewrites; random seeds, temperature, and tool versions must be frozen; and all artifacts must be written to immutable logs for easy recalculation. For enterprises, this mechanism naturally fits R&D efficiency scenarios: transforming requirements into formal acceptance conditions, optimization suggestions into executable patches, and letting CI, fuzz testing, and performance regression determine whether to merge. If a team needs to build such a pipeline, they can use a combination of general cloud capabilities—for instance, elastic computing for search, object storage for artifacts, container platforms for isolated execution, and log services with monitoring alerts to track budgets and anomalies. On Baidu AI Cloud, such a neutral pipeline can also be built using these general categories. The key remains verifier sovereignty and data governance, rather than any specific terminology.
Landing Boundaries and Human-AI Division of Labor
Verifiable closed loops excel at three types of problems: those with clear objective functions, enumerable or prunable solution spaces, and mechanically judgeable correctness. They are not suitable for directly answering “which direction is worth pursuing,” “is this result important,” or “does this failure imply a paradigm error.” Therefore, a reasonable division of labor is: humans propose valuable problems and set constraints and aesthetic standards; agents expand the search, generate candidates, and maintain proof drafts and counterexamples; the verification system guards correctness; and the review mechanism judges impact. For enterprise R&D, this means placing AI in the position of a “high-intensity candidate generator + strict gatekeeper,” rather than an unattended decision-maker.
Future improvement directions include stronger automatic lemma discovery, cross-task transfer, joint optimization of proofs and code, adversarial verifiers, and incorporating experimental design, data collection, and statistical analysis into the same ledger. What is truly worth noting is not whether one model “crushes” another, but a gradually clearing capability topographic map: within the verifiable domain, machines have begun to undertake substantive research labor; in the face of direction selection, value judgment, and major open problems, humans remain the final origin point.
