ProfileBench: Multi-Domain Benchmarking Framework

• ProfileBench is a multi-faceted benchmarking framework that decomposes performance metrics into profiles of latent cognitive and structural abilities across diverse domains.
• It employs controlled ablation techniques and gradient-based parameter importance scoring to diagnose model competence in LLMs, user profiling, optimization, and systems applications.
• By structuring evaluation around detailed profiles rather than single metrics, ProfileBench enhances interpretability and guides targeted improvements in model performance.

ProfileBench is most specifically the name of a mechanistic evaluation framework that decomposes LLM benchmark performance into mixtures of cognitively grounded abilities, rather than treating a benchmark as a proxy for a single capability (Kim et al., 23 Sep 2025). The same label, or an explicitly proposed shorthand, also appears in several adjacent literatures: industrial and streaming user profiling, profile reconstruction from personal albums, optimization benchmarking via performance profiles, and cross-application systems profiling (Li et al., 23 Sep 2025, Wang et al., 25 May 2026, Yan et al., 15 Jun 2026, Hekmati et al., 2018, Porcelli et al., 2017, Quackenbush et al., 2017, Prottasha et al., 15 Feb 2025). A plausible implication is that “ProfileBench” functions less as a single canonical benchmark than as a recurring design pattern: benchmarking by constructing, auditing, or exploiting profiles of latent structure.

1. Scope and naming

The label has been used in multiple, partly independent senses. In "Benchmark Profiling: Mechanistic Diagnosis of LLM Benchmarks" (Kim et al., 23 Sep 2025), ProfileBench is the author-introduced name of a framework for profiling benchmarks into ten abilities. In "Conf-Profile: A Confidence-Driven Reasoning Paradigm for Label-Free User Profiling" (Li et al., 23 Sep 2025), ProfileBench is the official name of an industrial benchmark for user profiling derived from Douyin. In "StreamProfileBench: A Benchmark for Fine-Grained User Profile Inference in Real-World Streaming Scenarios" (Wang et al., 25 May 2026), the name is extended to a streaming setting. In "PAL-Bench: Evidence-Grounded Profile Reconstruction from Longitudinal Personal Albums" (Yan et al., 15 Jun 2026), the benchmark is presented as a ProfileBench specialized to album-scale profile reconstruction. By contrast, in the optimization and systems papers below, “ProfileBench” is a convenient framework label used in the accompanying description rather than the paper’s official title (Hekmati et al., 2018, Porcelli et al., 2017, Quackenbush et al., 2017, Prottasha et al., 15 Feb 2025).

Usage	Official status in source	Domain
ProfileBench	Official	Mechanistic diagnosis of LLM benchmarks
ProfileBench	Official	Industrial user profiling
StreamProfileBench	Official	Streaming user profiling
PAL-Bench as a ProfileBench	Official framing	Album-scale profile reconstruction
“UserProfileConstruction + UserProfileUpdate” benchmark suite	Editorially proposed shorthand	Static and updating user profiles
ProfileBench for performance/data profiles	Editorial shorthand	Optimization and systems benchmarking

A common misconception is to treat the name as denoting one benchmark with one task. The sources instead describe several benchmarks and profiling frameworks that share an emphasis on latent structure, controlled evaluation, and profile-based diagnosis.

2. Mechanistic diagnosis of LLM benchmarks

In its most developed usage, ProfileBench is a diagnostic framework for benchmark audit and model interpretability. It addresses two stated problems: benchmarks are colloquially treated as if they “are” single capabilities, and this single-label view contributes to the performance–perception gap, because real benchmark success often depends on several skills in combination rather than one isolated faculty (Kim et al., 23 Sep 2025). The framework therefore constructs an ability profile for each benchmark by defining ten abilities grounded in Cattell–Horn–Carroll theory, identifying model parameters most important for each ability through gradient-based importance scoring, ablating those parameters, and measuring the normalized drop in benchmark performance.

The ten operationalized abilities are Analogical Reasoning, Commonsense Causal Reasoning, Contextual Recall, Deductive Reasoning, Inductive Reasoning, Long-Term Knowledge Recall, Quantitative Reasoning, Semantic Relationship Comprehension, Spatial Geometrical Reasoning, and Temporal Reasoning. Each ability has a dedicated diagnostic dataset $D_a$ consisting of 2000 4-choice MCQs, for 20,000 items in total. The items are designed to isolate the target skill, and expert review reports 92.2% correct label selection across a stratified sample. Operationalization is explicit: Analogical Reasoning uses proportional analogies with distractors that match surface similarity but break the relation; Deductive Reasoning uses first-order logic templates with distractors violating exactly one logical step; Spatial Geometrical Reasoning uses textual object-position and shape-relation descriptions with mirror or rotation distractors; Temporal Reasoning uses passages with explicit or implicit time markers and distractors that permute the true order.

Parameter importance is estimated with a first-order Taylor saliency approximation. For ability $a$ and parameter $\theta_j$, the importance score is

$$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$

Gradients are obtained by briefly fine-tuning on $D_a$ solely to estimate accurate gradients; the fine-tuned state is then discarded. Parameters are ranked by $I^{a}_{j}(\theta)$, and ablation is restricted to MLP-layer weights. The top $k$ percent, with $k = 1.024\%$, are selected per ability. The paper reports that a preliminary sweep over $k \in \{0.001\%, \ldots, 4.096\%\}$ found $k = 1.024\%$ to yield clear ability-specific signal with minimal collateral damage.

The central metric is the Ability Impact Score. For benchmark $a$0, baseline performance $a$1, ablated performance $a$2, and chance level $a$3, the score is

$a$4

AIS near 1 indicates strong reliance on ability $a$5, AIS near 0 weak reliance, and $a$6 negative transfer, meaning that ablating the ability improves performance. Chance-level normalization is benchmark-specific: 0.25 for 4-choice, 0.5 for 2-choice, 0.2 for 5-choice, and 0 for generation tasks.

The experimental setup profiles three instruction-tuned models—Llama-3.1-8B-Instruct, Qwen/Qwen2.5-7B-Instruct, and mistralai/Mistral-7B-Instruct-v0.3—across ten benchmarks: ARC-Challenge, HellaSwag, WinoGrande, CommonsenseQA, LogiQA, BIG-Bench Logical Deduction, GSM8K, Natural Questions Open, HumanEval, and MBPP. Evaluation uses the EleutherAI LM Evaluation Harness with each benchmark’s standard metric, greedy decoding, and BF16 execution.

The reported findings are fourfold. First, most benchmarks draw on several abilities rather than one. ARC-Challenge, often labeled “reasoning,” is centered on Analogical Reasoning with AIS $a$7 and Inductive Reasoning with AIS $a$8, while Long-Term Knowledge is low at $a$9. HellaSwag, intended as commonsense evaluation, shows a broad footprint with Inductive $\theta_j$0, Spatial $\theta_j$1, Contextual Recall $\theta_j$2, Analogical $\theta_j$3, and several others near $\theta_j$4–$\theta_j$5. GSM8K shows notable AIS in Long-Term Knowledge $\theta_j$6 and Semantic Relationship $\theta_j$7, with Quantitative Reasoning moderate at $\theta_j$8, indicating that math word-problem performance blends quantity, semantic parsing, and stored schemas.

Second, benchmarks with similar labels rely on distinct mixtures. CommonsenseQA peaks on Inductive $\theta_j$9 and also draws on Deductive $$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$0, Quantitative $$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$1, Spatial $$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$2, and Analogical $$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$3. NQ-Open emphasizes Temporal $$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$4, Spatial $$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$5, Semantic Relationship $$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$6, Long-Term Knowledge $$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$7, and Contextual Recall $$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$8. The paper links this divergence to sharply different accuracies: Llama-3.1-8B-Instruct scores 77.1% on CommonsenseQA but 17.9% on NQ-Open, while Qwen2.5-7B-Instruct scores 82.7% versus 4.7%.

Third, code benchmarks demand broad, multi-skill competence. HumanEval is driven most by Semantic Relationship $$I^{a}_{j}(\theta) \approx \left| \frac{\partial L(D_a, \theta)}{\partial \theta_j} \cdot \theta_j \right|.$$9, with Long-Term Knowledge $D_a$0, Temporal $D_a$1, Quantitative $D_a$2, Inductive $D_a$3, and Contextual Recall $D_a$4 also substantial. MBPP “lights up nearly every axis,” including Inductive $D_a$5, Spatial $D_a$6, Quantitative $D_a$7, Long-Term Knowledge $D_a$8, and Commonsense $D_a$9. The paper therefore argues that code-generation benchmarks reward broad, multi-skill improvement and show only modest gains from narrow domain-specific fine-tuning.

Fourth, abilities irrelevant to the task can harm performance. LogiQA has negative AIS for Long-Term Knowledge $I^{a}_{j}(\theta)$0, Quantitative $I^{a}_{j}(\theta)$1, and Semantic Relationship $I^{a}_{j}(\theta)$2. WinoGrande has negative AIS for Temporal $I^{a}_{j}(\theta)$3 and Semantic $I^{a}_{j}(\theta)$4. The reported increases are small but consistent, on the order of 1–2 percentage points, and are interpreted as cases in which external knowledge or heuristics lure the model toward plausible-but-invalid distractors.

Methodological validation emphasizes the ablation regime. MLP-only ablation is said to preserve fluency and targeted weakening, whereas all-layer ablation causes severe degradation and incoherence. Illustrative numbers are provided: GSM8K exact match drops from 0.773 to 0.7354 under MLP-only ablation of Contextual Recall but collapses to 0.1024 under all-layer ablation; NQ-Open exact match drops from 0.1789 to 0.1202 under MLP-only versus 0.0374 under all-layer ablation. Cross-model robustness is summarized by Jensen–Shannon Similarity on min–max normalized AIS distributions; all bars exceed 0.5, ranging from 0.53 to 0.89 with mean 0.64.

The framework’s stated limitations are also structural. The ten abilities reflect CHC-inspired choices rather than a uniquely correct taxonomy. Mild overlap between abilities is expected. Parameter superposition and nonlinear dynamics may cause misattribution when abilities co-recruit shared subcircuits. Synthetic diagnostic datasets may differ from real-world distributions. AIS can be inflated when baseline performance is very low because the denominator is improvement over chance, and generation tasks assume chance level 0. These caveats are integral to the interpretation of ProfileBench as a causal-ablation diagnostic rather than a complete theory of model competence.

3. User profiling benchmarks: industrial, open-source, and streaming variants

In user modeling, ProfileBench is the official name of an industrial benchmark built from a real-world, large-scale video platform, Douyin (Li et al., 23 Sep 2025). It is described as the first industry-level benchmark for user profiling grounded in a real-world platform, with heterogeneous inputs and a closed-set taxonomy over six core attributes. These attributes are Gender (3 labels), Age (7), Industry (14), Occupation (20), Life Stage (13), and Education Level (5), each including an explicit Unknown (NA) option. The data combine behavioral cues—white-box textual keyphrases summarizing watched, searched, or submitted video content—with optional, noisy demographic cues such as age, gender, region, and personal signature. The training corpus contains 20,000 unlabeled users, split into 10,000 for confidence-hinted data synthesis and supervised fine-tuning and 10,000 for confidence-guided unsupervised reinforcement learning; evaluation uses a 1,000-user human-annotated gold set.

The associated framework, Conf-Profile, is organized as a two-stage paradigm. A teacher LLM, DeepSeek-R1, generates per-attribute triplets $I^{a}_{j}(\theta)$5 consisting of tag, confidence on a 1–5 scale, and textual reasoning evidence. Parallel sampling with $I^{a}_{j}(\theta)$6 outputs is aggregated by confidence-weighted voting,

$I^{a}_{j}(\theta)$7

followed by equal-frequency confidence calibration into five bins per attribute. The student model is Qwen3-8B, trained first by structured distillation and then by confidence-guided GRPO-based unsupervised reinforcement learning. On ProfileBench, Qwen3-8B base attains average F1 59.06, Conf-Profile with SFT reaches 69.67, and Conf-Profile with SFT+RL reaches 73.03, an improvement of 13.97. Per-attribute gains are especially large for Life Stage (+26.49), Industry (+20.53), Occupation (+18.37), and Education Level (+17.65). The paper also reports that confidence-thresholding, which maps predictions with $I^{a}_{j}(\theta)$8 to NA, yields a controllable precision–recall trade-off.

A distinct, open-source benchmark suite for user profiling is introduced in "User Profile with LLMs: Construction, Updating, and Benchmarking" (Prottasha et al., 15 Feb 2025). The paper explicitly does not use the name “ProfileBench”; its official datasets are "UserProfileConstruction" and "UserProfileUpdate." Together they function as a benchmark for two tasks: constructing structured profiles from biographies and updating those profiles when new information arrives. The source is the Wiki People dataset with 42,786 entries and average biography length of approximately 1,896 characters. Attribute frequencies reported in the paper include Name 41,230, Profession 39,209, Birth Place 26,930, Likes 22,973, Hobbies 16,349, Dislikes 14,201, and Education 12,878. About 10% of generated profiles are manually checked as “Gold Data.” The updating dataset modifies about 30% of profiles to simulate evolving information and contradictions. Across Bloom-7B, Llama2-7B, Mistral-7B, Falcon-7B, and Gemma-7B, fine-tuning yields strong gains on both tasks, with Mistral-7B reaching F1 93.84 on construction and 95.08 on updating.

StreamProfileBench extends the profile-benchmark idea to continuous state maintenance under streaming user-generated content (Wang et al., 25 May 2026). The benchmark uses over 120,000 posts from 7,451 real users across five platforms—Sina Weibo, Xiaohongshu, Toutiao, Zhihu, and Douban—and formalizes streaming user profiling as

$I^{a}_{j}(\theta)$9

rather than a one-shot mapping from a static corpus. Evaluation is annotation-free: future user-authored interest anchors are extracted from the next batch and used as ground truth, while models must select exactly the positive-anchor budget from a candidate pool containing positives and structured distractors. The benchmark reports two-level macro-averaging, an Overall Recall main metric, and $k$0, a harmonic mean of novelty and stability recalls. The empirical headline is a systemic conservative bias: models retain past interests more easily than they detect emerging ones. Gemini-3-Flash attains the best reported macro Overall Recall at 52.26% and $k$1 at 54.97%, but even it shows Stability Recall 70.17 versus Novelty Recall 45.83. The paper argues that decay sensitivity, rather than coarse semantic understanding, is the persistent failure mode, since decay distractor errors are substantially higher than peer, viral, or random distractor errors.

Taken together, these user-profiling variants differ sharply in supervision regime, ontology, and temporal assumptions. The industrial benchmark emphasizes heterogeneous but consented data and label-free training; the open-source suite emphasizes biography-to-structure extraction and controlled profile updating; the streaming benchmark emphasizes persistent state, interest decay, and annotation-free temporal evaluation. This suggests that “ProfileBench” in user modeling names a family of evaluation problems rather than a single task definition.

4. Album-scale profile reconstruction and provenance-aware evaluation

PAL-Bench presents a different specialization: evidence-grounded reconstruction of profiles and social catalogs from longitudinal personal albums (Yan et al., 15 Jun 2026). The task is defined as data integration over weak-schema, multimodal records. Given only public, perception-derived album observations, an agent must discover owner facts, perform social identity binding, extract relations conditioned on correct identity binding, and cite provenance. Inputs are public albums $k$2 with records $k$3, where $k$4 includes release-safe metadata such as coarse location and device, and $k$5 includes caption, OCR strings, text entities, and remapped public face identifiers. Private owner profiles, social graphs, target definitions, evidence paths, and public–private identifier maps remain hidden.

The benchmark’s scale is explicit: 50 synthetic users, 36,659 released public photo records, 2,799 targets, 840 owner fact atoms, 653 social persons, 703 public face clusters, 2,947 planned events, and 4,869 key-evidence photos. Average album size is 733 photos per user, with range 563–1,000; average target count is 56 per user, with range 34–76. Coverage statistics include OCR in 82.4% of photos, text entities in 64.2%, and visible faces in 50.3%. Stress properties include key-evidence fraction 12.87%, distractor/key ratio 7.17×, mean evidence span 11.84 months, distinct months per fact 3.89, cross-modal owner paths 76.1%, and maximum face-cluster appearances 477.

PAL-Bench is built by an Evidence Compiler that synthesizes latent private worlds, programs target-level evidence paths, renders album pixels, re-measures them through perception pipelines, verifies whether public measurements realize private requirements, and exports audited public and private views. Construction outcomes are reported in detail: 38,378 manifest photos, 22,516 passed requirement checks, 1,183 were fixed via refinement, 14,636 were ambient with no requirement, 43 were unresolved or error, and 1,719 records were dropped at export, yielding 36,659 released public records across 50 of 50 users passing the audit.

Evaluation uses seven metrics. Owner Fact Recovery measures macro-averaged recovery of owner facts. Person Identity Recovery combines identity discovery and identity binding into an F1-style harmonic mean, with PIR-hard restricting to hard persons whose evidence paths require cross-event or cross-month alignment. PRR-ID measures relation or category accuracy conditioned on correct identity binding. Evidence Faithfulness Score averages fixed-judge sufficiency scores in $k$6 over matched claims with cited public photo IDs. Expected Calibration Error is computed over 10 equal-width confidence bins, and nLLM records the number of reconstruction LLM calls per user. The evaluation contract explicitly separates fact correctness, identity binding, and evidence sufficiency.

The benchmark’s principal empirical claim is a gap between plausible owner summarization and faithful social reconstruction. PAL-TRACE, the reference framework that freezes identity bindings before owner-fact mining, achieves the best reported macro averages among the main systems: OFR 0.6057, PIR 0.4792, PIR-hard 0.2684, PRR-ID 0.8384, EFS 0.3764, ECE 0.3236, and nLLM 7.44. Text-only and long-context baselines can reach nontrivial OFR while remaining weak on identity binding. Several baselines show high PRR-ID once binding is correct, indicating that relation labeling is easier than recurring-identity resolution. Compute-matched diagnostics further show that added call budget alone does not close the identity gap.

The benchmark also includes a privacy-preserving audit on real private albums with 10 volunteers. Reported findings include 10 of 10 albums spanning at least 2 years, median 72% non-public-shareable content, recurring non-owner people in all albums with median 11 identities, 101 of 120 fact chains requiring at least 2 photos, 79 of 120 requiring at least 2 modalities, and 94 of 120 relying on sensitive evidence. The conclusion drawn in the paper is that PAL-Bench’s controlled evidence structures occur in real albums, but equivalent real releases remain privacy-prohibitive.

5. Performance profiles, nested profiles, and profile-based training

In optimization benchmarking, “ProfileBench” names a methodological use of performance and data profiles rather than a single dataset (Hekmati et al., 2018, Porcelli et al., 2017). The foundational objects are the problem set $k$7, the solver set $k$8, and a performance measure $k$9, typically runtime or function evaluations. Standard performance profiles use the ratio

$k = 1.024\%$0

and define the profile

$k = 1.024\%$1

Here, $k = 1.024\%$2 is the fraction of problems on which solver $k = 1.024\%$3 is fastest, and $k = 1.024\%$4 is its success rate. The method is robust to outliers, but the 2018 paper identifies a specific flaw: rankings among non-best solvers can be biased and unstable because every solver is normalized against the best solver on each problem.

The paper "Nested Performance Profiles for Benchmarking Software" addresses this by iteratively removing the current best solver, recomputing ratios among the remaining solvers, and averaging the resulting “waves” of profiles (Hekmati et al., 2018). The overall nested performance profile is

$k = 1.024\%$5

with $k = 1.024\%$6 the number of waves. The paper’s Theorem 2 states that standard performance profiles are sensitive to elimination of the best solver, while nested performance profiles are not. It also states robustness to single-problem perturbations and an $k = 1.024\%$7-style continuity bound under small ratio perturbations. The artificial five-problem, three-solver example reproduced from Gould and Scott is central: on $k = 1.024\%$8, solver A appears dominant, but after removing A, solver B becomes best among $k = 1.024\%$9. Nested performance profiles recover the ordering $k \in \{0.001\%, \ldots, 4.096\%\}$0 without requiring such sequential replotting.

"A note on using performance and data profiles for training algorithms" extends the profile concept from evaluation to optimization of algorithm parameters (Porcelli et al., 2017). Data profiles define the fraction of problems solved within a normalized budget $k \in \{0.001\%, \ldots, 4.096\%\}$1, while performance-profile training maximizes improvement in the area under a performance profile relative to a baseline configuration. The paper uses area-under-profile objectives as training criteria for the BFO derivative-free optimizer on 55 small bound-constrained CUTEst problems. It reports four trained parameter vectors: average training with 17% gain, robust training with 7% gain, performance-profile training with 1% gain in the $k \in \{0.001\%, \ldots, 4.096\%\}$2 area objective, and data-profile training with 7% gain in the $k \in \{0.001\%, \ldots, 4.096\%\}$3 area objective. The qualitative conclusion is that profile-driven training improves the distribution of successes across the budgets or relative performance ratios emphasized by the chosen window, even when total-evaluation gains are larger under other objectives.

These works treat profiles not as metadata but as benchmarking primitives. A benchmark is the empirical distribution induced by solver behavior over a problem suite and, in the nested case, over successive comparison contexts. This is methodologically different from LLM or user-profile benchmarks, but the shared logic is still one of structured decomposition: instead of reducing software evaluation to a single scalar, profiles expose where and under what tolerances performance is achieved.

6. Cross-application transfer and recurrent design themes

A systems variant appears in "Beyond Profiling: Scaling Profiling Data Usage to Multiple Applications" (Quackenbush et al., 2017). The central question is whether profiling information from application $k \in \{0.001\%, \ldots, 4.096\%\}$4 on machine 1 can be used to make application $k \in \{0.001\%, \ldots, 4.096\%\}$5 run more efficiently on the same machine, thereby avoiding first-run profiling overhead for $k \in \{0.001\%, \ldots, 4.096\%\}$6. The proposed method has a training phase, which executes a corpus under multiple predefined configurations and learns a mapping from signatures to optimal configurations, and an execution phase, which continually collects signatures for unseen applications and reconfigures on the fly. The proof-of-concept target is last-level-cache block power-gating.

The paper uses Working Set Signatures for phase detection and Memory Access Signatures for configuration selection. MAS is a 64-bit shift register updated on every LLC access, with 1 for hit and 0 for miss. Two model classes are used: BLOOM, a set of Bloom filters with one filter per configuration, and ANN, a supervised feed-forward network with 64 inputs, 32 hidden units, and 5 outputs. The configuration space consists of five discrete LLC sizes: 20%, 40%, 60%, 80%, and 100% of the maximum LLC. The objective is to choose the smallest LLC size that does not increase execution time relative to the full cache.

The reported transfer results on PARSEC 2.1 benchmarks are concrete. BLOOM turns off an average of 19% of LLC blocks, ANN 5.4%, and EWSS 4.5%, while average miss-rate increase is approximately 2.8% versus full cache for BLOOM and ANN. Execution-time differences among algorithms are reported as less than 1%. For blackscholes with 4K data, average enabled LLC is about 80% for ANN, 61% for BLOOM, and 94% for EWSS. The paper interprets this as evidence that simple signatures can capture reusable hardware/software interaction patterns across applications on a fixed machine.

Across these otherwise disparate usages, several recurring design themes are visible. First, many ProfileBench-style systems reject one-number evaluation in favor of structured decomposition: AIS vectors in LLM benchmark diagnosis, per-attribute predictions with calibrated confidence in user profiling, owner/identity/relation splits with evidence scoring in PAL-Bench, and $k \in \{0.001\%, \ldots, 4.096\%\}$7- or $k \in \{0.001\%, \ldots, 4.096\%\}$8-indexed profile curves in optimization benchmarking. Second, several works explicitly separate public observations from hidden evaluator state, as in PAL-Bench’s public/private contract and industrial ProfileBench’s anonymized, user-authorized inputs. Third, a number of them operationalize robustness through controlled perturbation or alternative contexts: MLP-only versus all-layer ablation, nested elimination of best solvers, streaming update loops, or compute-matched diagnostics.

The main limitations also recur. Taxonomies are chosen rather than discovered, whether CHC-inspired abilities, closed-set user attributes, or anchor schemas. Synthetic or generated diagnostics may differ from real distributions. Comparison sets matter: standard performance profiles depend on the solver set $k \in \{0.001\%, \ldots, 4.096\%\}$9, and benchmark profiles depend on the selected abilities or candidate pools. Privacy constrains evaluation in precisely those domains where richly structured profile reconstruction would be most realistic. This suggests that ProfileBench, across its variants, is best understood as a class of rigorously structured benchmarking frameworks that make latent dependencies explicit, rather than as a universal benchmark with a fixed ontology or metric.