LLM as a Testset

• LLM as a Testset is the use of large language models to generate test inputs, oracle labels, and evaluative artifacts with clearly defined authority sources and adjudication mechanisms.
• It encompasses diverse construction paradigms such as authentic human artifacts, controlled synthetic generation, and self-consistent benchmarks using transformation graphs.
• Evaluation logic combines runtime assertions, compiler tests, and quantitative metrics to ensure reliability, mitigate bias, and validate benchmark effectiveness.

Searching arXiv for the cited papers to ground the article in current literature. “LLM as a Testset” denotes the use of LLMs to construct the inputs, oracle labels, or evaluative artifacts that make a benchmark or test suite operational. In the broadest formulation, the LLM does not merely generate candidate test cases; it can also supply expected outputs, pass/fail decisions, criteria-based grades, or executable checks. Recent work realizes this idea in markedly different ways: exam-derived legal benchmarks with statement-level grading, adversarial test inputs for algorithmic problems, compiler validation suites, psychometrically designed personality inventories, domain-specific judge sets, and automatically generated code-task benchmarks built from cycle-based consistency constraints (Mughal et al., 6 Jul 2026, Wind et al., 11 Feb 2026, Yang et al., 13 Jun 2025, Sollenberger et al., 29 Jul 2025, Farchi et al., 2024).

1. Conceptual scope and source of authority

A central distinction in the literature is between the source of authority, the oracle form, and the adjudication mechanism. The source of authority is the basis for trusting a verdict; oracle form is the representation of the verdict, such as an assertion or expected value; adjudication mechanism is how pass/fail or a score is computed. This separation matters because labels such as “LLM-as-a-judge” describe a mechanism rather than a basis for trust. The systematic review on LLM-based test oracles identifies seven sources of authority, and reports that specification-derived authority is the largest single category, while 26 of 54 studies reach verdicts with no specification at all (Mughal et al., 6 Jul 2026).

Source of authority	Studies	Characterization
Specification-derived	28 of 54	External requirements, docs, API specs, BDD scenarios
Implementation-derived	11 of 54	Current code or existing tests as de facto behavior
Reference-differential	4 of 54	Independent implementation or consensus reference
Model-parametric	9 of 54	LLM prior used directly as oracle
Human-elicited	2 of 54	Human rubric or checklist applied by an LLM
Regression-from-prior-version	0 primary	Earlier version as ground truth not primary in corpus
Implicit-intrinsic signals	0 primary	Crashes or well-formedness only secondary signals

The same review reports that oracle forms are led by assertions (31), expected-output values (13), metamorphic relations (11), classifier or judge labels (10), properties (7), invariants (4), and exception oracles (1). Adjudication mechanisms are led by runtime assertion execution (33), exact or structural match (18), LLM-as-judge (15), human review (7), statistical or distributional checks (5), semantic similarity (4), and formal verification (2). In 41 of 54 studies the LLM generates an oracle artifact that is checked by another mechanism; in only 13 does the LLM serve directly as the oracle or judge (Mughal et al., 6 Jul 2026).

A related formulation is “LLM-as-a-benchmark,” in which an LLM both creates the test set and evaluates outputs. Translation benchmarking makes this composition explicit by separating testset-origin bias, evaluator-origin bias, and their combination, showing that benchmark generation and benchmark judgment are distinct components even when instantiated by the same model (Xu et al., 30 Sep 2025).

2. Construction paradigms

One major construction paradigm uses authentic human assessment artifacts as the substrate. SteuerEx is generated from authentic German university tax law examinations and encodes 115 examination questions, 1,035.5 total achievable points, and 752 independently scorable statements with varying weights. LLMzSzŁ likewise turns Polish national exams into a test-only benchmark: 4 exam types, 154 domains, and almost 19k closed-ended questions extracted from the Polish Central Examination Board. In both cases, authenticity, dated provenance, and structured answer keys are used to improve realism and contamination awareness (Wind et al., 11 Feb 2026, Jassem et al., 4 Jan 2025).

A second paradigm uses controlled synthetic generation with explicit structural constraints. In directive-based compiler testing, LLM4VV extracts several thousand valid directive–clause combinations from OpenMP and OpenACC V&V suites, then prompts a generative LLM to produce model-conformant tests in C, C++, and Fortran. TestCase-Eval treats the LLM as an automated test-set designer for algorithmic problems: given only a Codeforces problem statement, the model must generate $N$ test inputs that maximize coverage of incorrect human submissions, or a single targeted input that exposes a specific faulty implementation. ULT constructs 3,909 function-level Python tasks from The Stack v2, filtered to cyclomatic complexity $\ge 10$, verified for testability, and explicitly decontaminated by searching for leaked tests (Sollenberger et al., 29 Jul 2025, Yang et al., 13 Jun 2025, Huang et al., 1 Aug 2025).

A third paradigm builds self-consistent benchmarks from transformation graphs. In automatically generated code-task benchmarks, artifacts such as natural-language descriptions, code in multiple languages, and summaries are vertices in a graph $G=(V,E,L)$; cycles such as description $\rightarrow$ code $\rightarrow$ summary $\rightarrow$ description generate expectations like endpoint equivalence or cross-language summary consistency. This yields labels without requiring manual gold annotation for each item, because cycle-based invariants serve as automatically generated claims (Farchi et al., 2024).

A fourth paradigm emphasizes domain-balanced curation for LLM judging. Domain-specific evaluation sets for LLM-as-a-judge are built from manual category seeds, multilingual embeddings, k-NN labeling with entropy gating, and stratified sampling. The resulting evaluation set contains 1,573 samples across 14 categories covering finance, law, medical, math, coding, and eight languages, with roughly 90–100 prompts per category to reduce variability and tighten confidence intervals (Raju et al., 2024).

3. Evaluation logic and quantitative measures

The literature does not use a single metric family; instead, evaluation tracks the objective that the testset is supposed to measure. In exam-derived legal reasoning, SteuerEx scores free-form answers by decomposing reference solutions into “pointable statements” with maximum points $m_{q,i}$ and awarding statement-level scores $a_{q,i}$ with $0 \le a_{q,i} \le m_{q,i}$. Question and benchmark totals are then aggregated as

$$A_q = \sum_{i=1}^{n_q} a_{q,i}, \qquad M_q = \sum_{i=1}^{n_q} m_{q,i},$$

and

$\ge 10$0

This framework makes citation accuracy, subsumption steps, and numeric sub-results individually scorable, so that missing a statute citation or making an arithmetic error can reduce only the corresponding weighted statements rather than collapsing the entire answer (Wind et al., 11 Feb 2026).

In algorithmic test generation, TestCase-Eval distinguishes Fault Coverage and Fault Exposure. If $\ge 10$1 are generated tests and $\ge 10$2 is the set of incorrect submissions detected by test $\ge 10$3, then

$\ge 10$4

For targeted exposure, the model receives the faulty code and must craft a single input $\ge 10$5; the exposure rate is the average of $\ge 10$6 over sampled faulty implementations. This explicitly separates untargeted breadth from targeted adversarial precision (Yang et al., 13 Jun 2025).

Compiler-test generation uses a combined generative and discriminative metric suite. Generative quality is measured by Compilation Rate, Returned-0 Rate, and Pass@1; discriminative judging is measured by Accuracy, Normalized Bias, Permissiveness, Precision, Recall, F1-Score, and Matthews Correlation Coefficient. In unit test generation, ULT evaluates Pass@k, statement coverage, branch coverage, and mutation score, reporting on ULT averages of 41.32% accuracy, 45.10% statement coverage, 30.22% branch coverage, and 40.21% mutation score, all substantially below TestEval and PLT (Sollenberger et al., 29 Jul 2025, Huang et al., 1 Aug 2025).

Judge-oriented benchmarks add another layer of evaluation. Domain-specific evaluation-set construction uses Bradley–Terry aggregation with bootstrapped confidence intervals, then reports separability, agreement with Chatbot Arena, Spearman correlation, and pairwise Brier score. The reported benchmark reaches 84.44% separability, 84.44% agreement with Chatbot Arena, Spearman 0.915, and Brier 0.0417. Quantitative LLM judges post-hoc calibrate a base judge’s textual rationale and raw score through generalized linear models, improving MSE, MAE, classification accuracy, or pairwise ranking metrics depending on whether the task is absolute or relative (Raju et al., 2024, Sahoo et al., 3 Jun 2025).

4. Representative application domains

Legal reasoning provides one of the clearest high-fidelity realizations of LLM as a testset. SteuerEx mirrors real university grading in German tax law through weighted statements, partial credit, exact statutory citations, and numerical accuracy checks. It is explicitly intended to assess whether an LLM’s free-form answers meet academic and professional standards under rigid rubrics. A related high-stakes legal study uses the Polish National Appeal Chamber qualifying exam as a direct probe of both LLM-as-candidate and LLM-as-a-judge. There, models achieved strong formal extraction scores but failed the practical written judgment, and the LLM judge substantially inflated substantive scores relative to the official committee (Wind et al., 11 Feb 2026, Karp et al., 6 Nov 2025).

Software testing is another major application. LLM4VV uses a dual-LLM workflow in which one model generates directive-based compiler tests and another judges validity, with compile-and-run gating before LLM judgment to improve trust and reduce hallucinations. ULT and PLT instead evaluate the ability of LLMs to generate unit tests for real-world Python functions under contamination control and high structural complexity. TestCase-Eval shifts attention from code synthesis to test synthesis by measuring how well LLMs can design inputs that catch real incorrect competitive-programming submissions (Sollenberger et al., 29 Jul 2025, Huang et al., 1 Aug 2025, Yang et al., 13 Jun 2025).

Behavioral and psychometric evaluation show that “testset” need not mean executable code. TRAIT constructs an approximately 8,000-item multi-choice personality benchmark for LLMs from psychometrically validated instruments and ATOMIC10× scenarios, replacing introspective self-report with behavior-observation through choices in realistic situations. It reports lower refusal rate and lower sensitivity to prompt, option order, and paraphrase than earlier personality tests for LLMs (Lee et al., 2024).

Machine translation exposes another use case and a cautionary one. Here the LLM generates 200 source texts per language direction under controlled templates, and systems are evaluated by a reference-free metric or an LLM judge. The benchmark is useful for scalable ranking, but into-English low-resource directions show systematic self-bias when the generating model also participates in evaluation, especially when source diversity is low (Xu et al., 30 Sep 2025).

5. Reliability, bias, and statistical justification

A recurring controversy is whether LLM-produced labels can substitute for human labels. The Alternative Annotator Test addresses this directly by comparing the LLM’s alignment with a leave-one-out set of annotators to each human annotator’s own alignment with that same set. For annotator $\ge 10$7, the null hypothesis is

$\ge 10$8

with Benjamini–Yekutieli correction across annotators and a decision rule based on the winning rate

$\ge 10$9

If $G=(V,E,L)$0, the LLM is justified as replacing a majority of annotators. The same work introduces Average Advantage Probability,

$G=(V,E,L)$1

as an interpretable measure of the probability that the LLM’s annotation is at least as good as a randomly chosen human annotator’s annotation (Calderon et al., 19 Jan 2025).

A different reliability strategy calibrates the judge rather than replacing humans outright. Quantitative LLM judges keep the base LLM frozen and learn a lightweight regression or classification layer over the judge’s rationale embedding and score or preference probability. The framework instantiates least-squares, multinomial, Bradley–Terry–Luce, and two-headed BTL variants, and reports improvements over raw base judges on rating and pairwise tasks while remaining much cheaper than supervised fine-tuning (Sahoo et al., 3 Jun 2025).

Empirical alignment, however, is domain-sensitive. In SteuerEx, expert inter-rater reliability at the statement level is low, with $G=(V,E,L)$2, yet the automated evaluator aligns strongly with averaged human scores, with Kendall’s $G=(V,E,L)$3 and model-stratified $G=(V,E,L)$4–$G=(V,E,L)$5. By contrast, the Polish KIO exam study reports absolute pass/fail disagreement between the committee and the LLM judge: committee 0/3 pass versus LLM judge 3/3 pass, yielding Cohen’s $G=(V,E,L)$6. This suggests that LLM judging may be comparatively reliable for structured comparative benchmarking, while becoming unreliable in high-stakes legal grading where cardinal doctrinal errors matter more than surface coherence (Wind et al., 11 Feb 2026, Karp et al., 6 Nov 2025).

Self-bias is another fundamental reliability issue. Translation benchmarking formally decomposes self-bias into testset-origin bias, evaluator-origin bias, and combined benchmark bias, and reports that the combined setting amplifies bias relative to either component alone. The same study attributes part of the effect to low diversity in generated source texts and shows that selecting higher-diversity subsets reduces bias magnitudes (Xu et al., 30 Sep 2025).

6. Design principles and research agenda

Several design principles recur across the literature. First, the source of authority should be chosen explicitly: specification-derived authority is preferable when trustworthy requirements exist, implementation-derived authority is useful for regression, reference-differential designs are strong when independent implementations are available, and model-parametric authority is a last resort for interactive or visual domains with scarce executable specifications. Second, deterministic or executable adjudication should be preferred whenever possible: compile–run gating, runtime assertion execution, exact or structural match, and formal verification increase trust relative to unconstrained LLM judgment (Mughal et al., 6 Jul 2026, Sollenberger et al., 2024).

Third, high-quality domain testsets tend to share a common construction pattern: source authentic materials, decompose answers into atomic scorable units, separate citations and numerical sub-results into independent graded items, use machine-readable formats, and release score aggregation tooling. SteuerEx makes these design choices explicit, while domain-specific evaluation-set construction adds multilingual embeddings, k-NN labeling, entropy gating with $G=(V,E,L)$7 sent to a general category, and stratified sampling to maintain near-balanced category sizes (Wind et al., 11 Feb 2026, Raju et al., 2024).

Fourth, contamination and memorization require active control. LLMzSzŁ uses time stamps to support contamination-aware slices based on publication dates; ULT removes functions whose tests appear leaked in The Stack v2 and withholds ground-truth tests from release; PLT is retained specifically as a contrast set for memorization analysis. These practices indicate that an LLM-generated or LLM-mediated testset is not automatically informative unless leakage is measured or bounded (Jassem et al., 4 Jan 2025, Huang et al., 1 Aug 2025).

The research agenda remains substantial. The oracle review notes that 59% of studies rely on custom benchmarks, mutation analysis is rare despite being a gold standard for fault detection, and two authority sources—regression-from-prior-version and implicit-intrinsic signals—remain unoccupied as primary LLM contributions. A plausible implication is that the field is still converging on trustworthy benchmark design rather than merely scaling benchmark generation. The strongest recent results therefore come not from treating LLM output as self-authenticating, but from coupling LLM generation with explicit authority sources, structured scoring, statistical validation, and auditable execution pipelines (Mughal et al., 6 Jul 2026).