LLM as a Benchmark: Roles & Evaluation

• LLM as a Benchmark is a framework where LLMs act as evaluation subjects, judges, generators, and auditors in diverse applications.
• It assesses model performance through agentic tasks like spatial reasoning and legal analysis, reporting quantifiable improvements such as chain-of-thought gains.
• LLM-based benchmarking innovates by automating data generation and audit processes while addressing challenges in contamination, fairness, and reliability.

Across recent work, “LLM as a Benchmark” is best understood as a family of benchmark practices rather than a single protocol. In this usage, LLMs are benchmark subjects in increasingly agentic or domain-specific environments; they are also benchmark instruments, acting as judges, labelers, critics, and generators of evaluation data; and, in a further reflexive turn, model behavior is used to evaluate benchmark quality itself. The resulting literature spans spatially grounded building tasks, document reading systems, legal reasoning, public-sector evaluation, multilingual meta-evaluation, synthetic benchmark generation, and benchmark-auditing frameworks (Madge et al., 2024, Yuan et al., 2 Feb 2025, Zhou et al., 21 Apr 2025, Qian et al., 7 Jan 2026).

1. Conceptual scope

A useful synthesis is that papers assign LLMs four distinct benchmark roles. First, an LLM can be the object being benchmarked, as in spatial instruction following, document reading, legal reasoning, or agentic decision making. Second, an LLM can be the judge that scores other models or trajectories. Third, an LLM can be the generator that creates benchmark items, labels, or synthetic environments. Fourth, model outputs can be used to audit the benchmark itself, asking whether a benchmark is discriminative, contamination-resistant, or aligned with expected capability orderings.

Role	Representative formulation	Representative papers
Benchmark subject	LLM performance on task suites, environments, or workflows	(Madge et al., 2024, Zou et al., 2024, Fan et al., 19 May 2025, Wang et al., 9 Jan 2026)
Judge or evaluator	LLM-as-a-Judge, reward-model meta-evaluation, trajectory scoring	(Tang et al., 20 May 2026, Son et al., 2024, Zhou et al., 21 Apr 2025, Shao et al., 5 Aug 2025)
Generator or labeler	Automated benchmark construction, synthetic benchmark generation	(Yuan et al., 2 Feb 2025, Wang et al., 25 Jun 2025, Farchi et al., 2024)
Benchmark auditor	Evaluation of benchmark quality or contamination resistance	(Qian et al., 7 Jan 2026, Al-Lawati et al., 19 May 2026)

This taxonomy suggests that “benchmark” no longer refers only to a static test set with fixed labels. In the newer literature, benchmarking is increasingly procedural: it may involve interactive environments, pairwise comparisons, synthetic data generation, judge calibration, or even benchmark compression and latent-space release. A plausible implication is that benchmark design is becoming part of model-systems engineering rather than a purely downstream measurement step.

2. LLMs as benchmark subjects in task-grounded evaluation

One major strand uses benchmarks to test whether LLMs can function as components in realistic systems rather than as text-only answerers. The Minecraft builder benchmark adapts the Minecraft Builder Dialog Agent task into three synthetic task types—absolute addressing, relative addressing, and primitive shapes—to probe spatial reasoning and vector-based math. On Llama-3-70b-Instruct, chain-of-thought improved absolute addressing from 42.98 to 76.5 and relative addressing from 82.02 to 95.8, while primitive shapes changed only from 59.02 to 60.3. The same work highlights recurring failures in coordinate consistency, including confusion about handedness despite the explicit convention “$Z$ positive for south” (Madge et al., 2024).

A closely related shift appears in full-system document evaluation. DocBench evaluates LLM-based document reading systems on raw PDF files rather than plain text, requiring file parsing, layout understanding, metadata extraction, multimodal content handling, and long-context reasoning. It contains 229 real documents and 1,102 questions across Academia, Finance, Government, Laws, and News. Human annotators reach 81.2% overall accuracy, while the best evaluated system reaches 69.8%; on unanswerable questions, humans score 82.2% while GPT-4 in the file-upload setting scores 37.1%, exposing a substantial groundedness gap (Zou et al., 2024).

The same expansion from static QA to workflow competence appears in legal and agentic settings. LEXam derives 4,886 law exam questions from 340 law exams spanning 116 courses and evaluates both multiple-choice and open-ended legal reasoning; LegalAgentBench evaluates Chinese legal agents over 17 corpora, 37 tools, and 300 tasks, measuring both success rate and progress rate; TowerMind evaluates LLM agents in a tower defense environment with five benchmark levels, multimodal observations, and hallucination measured by valid action rate; and EconEvals benchmarks agents that act over 100 periods in unknown economic environments such as procurement, scheduling, and pricing (Fan et al., 19 May 2025, Li et al., 2024, Wang et al., 9 Jan 2026, Fish et al., 24 Mar 2025).

Domain-specific benchmarking also extends beyond task performance into governance and population-sensitive risk. Safe-Child-LLM evaluates 200 adversarial prompts split evenly across ages 7–12 and 13–17, using both binary harmfulness labels and a 0–5 action taxonomy ranging from “Strong Refusal” to “Support or Intensification.” MÖVE evaluates 39 models for the German public sector across performance criteria—summarization, question answering, and topic extraction—and governance criteria including hallucination tendencies, energy consumption, provider transparency, and alignment with German constitutional values (Jiao et al., 16 Jun 2025, Dalerci et al., 11 Jun 2026).

Taken together, these benchmarks redefine the benchmark subject. The model is increasingly tested as a spatial reasoner, document reader, legal assistant, tool-using agent, public-sector component, or safety-critical interlocutor, rather than as a generic next-token predictor.

3. LLMs as judges, critics, and evaluators

A second strand places LLMs inside the benchmark as evaluators. RankJudge is exemplary: it generates paired multi-turn conversations grounded in the same reference document, with exactly one injected flaw in exactly one turn. The judge must jointly predict which conversation is better, which turn is flawed, and what kind of flaw it is, under the strict criterion

$$c=\mathbf{1}\!\left[\hat{y}=y^\star \wedge \hat{r}=r^\star \wedge \hat{t}=t^\star\right]\in\{0,1\}.$$

This moves beyond verdict-only judging toward diagnostic understanding (Tang et al., 20 May 2026).

Other work evaluates judge models more directly. MM-Eval is a multilingual meta-evaluation benchmark for LLM-as-a-Judge and reward models, comprising five core subsets across 18 languages and a Language Resource subset spanning 122 languages. It measures not only pairwise ranking accuracy but also score consistency and fairness across languages. The reported overall average accuracy is 68.9%, and performance on Linguistics and Safety is 12.8% and 18.4% lower, respectively, in low-resource languages than in English, indicating that evaluator quality degrades under multilingual resource imbalance (Son et al., 2024).

JETTS asks a narrower systems question: whether LLM judges help test-time scaling. Across math reasoning, code generation, and instruction following, it evaluates judges in response reranking, step-level beam search, and critique-based response refinement. Helpfulness in reranking is normalized as

$$h = \frac{p_\mathrm{judge} - p_\mathrm{greedy}}{p_\mathrm{oracle} - p_\mathrm{greedy}}.$$

The main finding is asymmetric: judges are competitive with outcome reward models in reranking, consistently worse than process reward models in beam search, and their natural-language critiques are currently ineffective for refinement (Zhou et al., 21 Apr 2025).

Trajectory-level judging appears in cybersecurity and legal reasoning as well. CTFJudge compares expert writeups with agent traces along six dimensions—vulnerability understanding, reconnaissance thoroughness, exploitation methodology, technical accuracy, efficiency of approach, and adaptability—and summarizes them with the CTF Competency Index,

$$\mathrm{CCI}(T, G) = \sum_{i=1}^{n} w_i\,F_i(T, G), \qquad \sum_{i=1}^{n} w_i = 1.$$

LEXam uses GPT-4o as a judge for open-ended legal answers and validates the setup with three human legal experts, reporting a winning rate $\omega = 0.67$ in the Alternative Annotator Test, above the recommended threshold $\omega \ge 0.5$ (Shao et al., 5 Aug 2025, Fan et al., 19 May 2025).

The literature is, however, not uniformly optimistic about LLM judging. Creativity Benchmark, built from 100 brands and 11,012 anonymised pairwise comparisons from 678 practising creatives, finds weak and unstable judge-human alignment: Spearman correlations are prompt- and judge-dependent, often around 0.2–0.35, and automated judges cannot substitute for expert human evaluation in brand-constrained creativity (Bhat et al., 5 Sep 2025).

4. LLMs as benchmark generators and labelers

A third strand treats LLMs as infrastructure for benchmark production. The enterprise benchmark uses a 14-task framework grounded in Bloom’s Taxonomy and a scalable pipeline that combines LLM-as-a-Labeler, LLM-as-a-Judge, corrective retrieval-augmented generation (CRAG), and human validation. The benchmark is built from Confluence, Rovo chat, customer feedback, Slack query data, and developer documentation, and contains about 9,700 samples. This is a direct attempt to turn benchmark construction into an enterprise workflow rather than a fully manual annotation exercise (Wang et al., 25 Jun 2025).

BenchMaker pushes the same idea toward generic benchmark factories. It proposes an automated evaluation framework with four dimensions—Credibility, Diversity, Difficulty, and Benchmark-level quality—and ten criteria, then augments direct LLM prompting with stepwise self-correction, conflict-guided contrastive discrimination, difficulty diffusion, difficulty strategy guidance, AttrPrompt, and in-batch diversity boosting. Across multiple tasks, it reports highly consistent evaluation results across 12 LLMs, including 0.967 Pearson correlation against MMLU-Pro, at about $0.005 and 0.38 minutes per sample (Yuan et al., 2 Feb 2025).

For code tasks, automatic benchmark generation becomes more structural. One paper defines a multigraph

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

where vertices are artifact types and edges represent possible generations under either a trusted generator or a tested generator. Cycles in $G$ produce self-consistency expectations—for example, translating code across languages and back, or moving from description to code to summary to description—so that benchmark instances and judge labels can be derived automatically (Farchi et al., 2024).

Executable specifications form another variant of LLM-generated benchmarking. WebApp1K contains 1,000 test-driven development tasks across 20 application domains, where tests are both prompt and verification. The benchmark reports pass@1, pass@5, and pass@10, and the central conclusion is that instruction following and in-context learning are more critical for TDD success than general coding proficiency or pretraining knowledge (Cui, 13 May 2025).

This body of work suggests that benchmark generation is being reconceived as a controlled LLM pipeline with explicit validation stages, rather than a one-time dataset release.

5. Benchmarking the benchmark: quality, compression, and contamination resistance

Once LLMs are used to build and score benchmarks, benchmark quality itself becomes an object of measurement. Benchmark^2 formalizes this with three metrics. Cross-Benchmark Ranking Consistency is

$$\text{CBRC}(B_i) = \frac{1}{n-1} \sum_{j \neq i} \tau(r_i, r_j),$$

where $\tau$ is Kendall’s $$c=\mathbf{1}\!\left[\hat{y}=y^\star \wedge \hat{r}=r^\star \wedge \hat{t}=t^\star\right]\in\{0,1\}.$$0 over model rankings. Capability Alignment Deviation penalizes within-family inversions through

$$c=\mathbf{1}\!\left[\hat{y}=y^\star \wedge \hat{r}=r^\star \wedge \hat{t}=t^\star\right]\in\{0,1\}.$$1

Across 15 benchmarks and 11 LLMs from four model families, the paper reports substantial quality variation and shows that selective benchmark construction using about 35% of the original instances can retain average Kendall’s $$c=\mathbf{1}\!\left[\hat{y}=y^\star \wedge \hat{r}=r^\star \wedge \hat{t}=t^\star\right]\in\{0,1\}.$$2 with the full benchmark (Qian et al., 7 Jan 2026).

A different benchmark-level concern is contamination. The contamination-resistance paper argues that benchmark datasets should be “unlearnable-but-inferable”: models should be able to use them for inference, but gradient-based training on the released form should fail to improve generalization. The proposed release format replaces plaintext with latent representations, specifically the KV cache $$c=\mathbf{1}\!\left[\hat{y}=y^\star \wedge \hat{r}=r^\star \wedge \hat{t}=t^\star\right]\in\{0,1\}.$$3 and the penultimate-layer hidden state $$c=\mathbf{1}\!\left[\hat{y}=y^\star \wedge \hat{r}=r^\star \wedge \hat{t}=t^\star\right]\in\{0,1\}.$$4, together with the plaintext answer or label. This reframes the benchmark as a latent evaluation object rather than a public text corpus (Al-Lawati et al., 19 May 2026).

Both lines of work alter the conventional ontology of a benchmark. A benchmark is no longer merely a held-out set; it may be a statistically audited subset, a latent-space release, or an adaptive construction whose quality is itself measured against peer benchmarks and family hierarchies.

6. Limitations, controversies, and emerging consensus

Several limitations recur across the literature. First, LLM judges remain useful but imperfect. MÖVE reports inter-judge agreement values ranging from Krippendorff’s $$c=\mathbf{1}\!\left[\hat{y}=y^\star \wedge \hat{r}=r^\star \wedge \hat{t}=t^\star\right]\in\{0,1\}.$$5 for Noise Sensitivity to $$c=\mathbf{1}\!\left[\hat{y}=y^\star \wedge \hat{r}=r^\star \wedge \hat{t}=t^\star\right]\in\{0,1\}.$$6 for QA Factual Correctness, and explicitly notes that inter-judge disagreement is larger than within-judge randomness. This indicates that judge choice is itself a benchmark design variable (Dalerci et al., 11 Jun 2026).

Second, automated evaluation is uneven across domains. Creativity Benchmark finds tightly clustered model performance, with top-bottom spread $$c=\mathbf{1}\!\left[\hat{y}=y^\star \wedge \hat{r}=r^\star \wedge \hat{t}=t^\star\right]\in\{0,1\}.$$7, implying a head-to-head win probability of only about 0.61; in that setting, different judges and rubrics can reorder models, and expert human evaluation remains necessary. JETTS reaches a related conclusion from another angle: natural-language critiques, though distinctive to judge models, do not yet provide reliable control signals for response refinement (Bhat et al., 5 Sep 2025, Zhou et al., 21 Apr 2025).

Third, multilingual and demographic fairness remain unresolved. MM-Eval shows that evaluator models assign lower scores to good responses and higher scores to bad ones in lower-resourced languages, compressing judgments toward the middle. Safe-Child-LLM shows that even strong frontier models can fail in child-facing safety scenarios, particularly when harmful intent is phrased mildly or innocuously. These findings indicate that an LLM benchmark can itself encode uneven evaluator behavior unless fairness is measured directly (Son et al., 2024, Jiao et al., 16 Jun 2025).

Fourth, synthetic and automated benchmarks trade annotation cost for new validation burdens. BenchMaker reports strong effectiveness and efficiency, but also documents a 7.5% error rate in manual inspection of 80 random samples. The enterprise benchmark retains human validation for low-confidence samples, and RankJudge drops the top 5% of pairs by BT Elo because the hardest tail concentrates ambiguity and noise (Yuan et al., 2 Feb 2025, Wang et al., 25 Jun 2025, Tang et al., 20 May 2026).

A broad consensus nonetheless emerges. LLM-based benchmarking is no longer confined to generic academic test suites. It increasingly measures grounded action, long-horizon behavior, domain fidelity, judge reliability, and benchmark quality itself. This suggests that “LLM as a Benchmark” names a shift from static accuracy measurement toward a layered evaluation stack in which LLMs are simultaneously benchmarked, used to benchmark, and used to diagnose the validity of the benchmarking process.