Silicon Bureaucracy and AI Test-Oriented Education: Contamination Sensitivity and Score Confidence in LLM Benchmarks
The paper critiques the institutionalization of LLM benchmarks as "Silicon Bureaucracy" and "AI Test-Oriented Education", arguing high scores often conflate exam-oriented competence with genuine generalization due to data contamination. It proposes an audit framework using a router-worker setup: clean-control routers transmit full questions while noisy routers delete, rewrite, and perturb before aggregation. For clean benchmarks, noisy aggregation should not systematically exceed the baseline; persistent above-baseline gains suggest contamination-related memory activation. The core finding—that 10 of 12 models exceed clean baselines under multi-router noisy conditions—challenges the interpretability of raw benchmark scores.
The paper makes a timely conceptual contribution by framing benchmark evaluation as institutionalized examination rather than neutral measurement. The router-worker audit framework is methodologically novel, and the empirical demonstration of heterogeneous contamination sensitivity across models (10/12 showing above-baseline gains at r=8) provides concrete evidence for long-suspected contamination issues. However, the theoretical framework conflates two distinct failure modes: benchmark leakage versus genuine capability for reconstructing information from noisy fragments. The claim that "deleted, rewritten, and perturbed fragments may be recombined into semantic-neighbor cues" (Section 5.1) is plausible but not directly validated—alternative explanations involving ensemble-like error correction or prompt diversity effects are not rigorously excluded.
The conceptual distinction between "exam-oriented competence" and "principled understanding" (Section 1) is well-articulated and aligns with construct validity critiques in the literature. The empirical heterogeneity finding is robust: Qwen3-Next-80B violates baseline in all 9 router settings while DeepSeek-Chat violates only once, demonstrating that "similar benchmark scores may carry substantially different levels of confidence" (Abstract). The directional transition hypothesis (H3) is supported by question-level analysis showing improve transitions (wrong→correct) rising from 112 to 180 while degrade transitions fall from 150 to 116 as router count increases from 1 to 9 (Section 5.3), suggesting systematic cue aggregation rather than random fluctuation.
Sample size is critically small: only n=100 questions drawn from an unspecified public benchmark with seed 42 (Section 4.1). This limits statistical power and generalizability. The theoretical justification for the clean baseline as "full-information transmission" is underdeveloped—if the single clean router suffers from position bias, output format sensitivity, or overthinking, the noisy conditions might legitimately improve performance through beneficial perturbation rather than contamination activation. The paper acknowledges $\varepsilon_m(q,\theta)$ should be small (Equation 13) but provides no empirical verification. Most critically, the mechanism linking above-baseline gains to "contamination-related memory traces" (Section 3.3) is inferred rather than demonstrated; the study lacks a validation set of confirmed-clean questions where the theory predicts no gains, nor does it compare against models with verified contamination-free training.
The evidence supports the existence of above-baseline anomalies but overstates the case for contamination as the exclusive mechanism. The comparison to Dekoninck et al. (2024) on ConStat is apt—the paper extends performance-based detection to multi-router aggregation—but the authors do not demonstrate that their method detects contamination that ConStat would miss. The discussion of Sun et al. (2025) on mitigation strategies correctly notes that "exact deduplication does not imply the disappearance of contamination" (Section 2.2), though the present work does not empirically validate this by testing models trained with vs. without deduplication. The heterogeneous sensitivity finding is novel compared to uniform contamination assumptions, but the comparison lacks baseline data showing that models with known clean training (e.g., specifically held-out benchmarks) exhibit no gains under the same protocol.
Reproducibility is severely limited. The paper mentions "code and data are available at" in some references but does not commit to releasing the router-worker implementation, prompts, or aggregation logic. The benchmark is vaguely identified only as "a public benchmark consisting of multiple-choice test questions" (Section 4.1)—this could be MMLU, ARC, or other, but is not specified, preventing exact replication. Critical hyperparameters (temperature, top-p, system prompts for routers) are not reported. Several models evaluated (Qwen3-Next-80B, Seed-2.0-Lite, Qwen3.5-35B) appear to be unreleased or proprietary, making independent verification impossible. The random seed 42 for question selection is provided, but without the scaffold code and exact API versions, the experiment cannot be reproduced.
Public benchmarks increasingly govern how large language models (LLMs) are ranked, selected, and deployed. We frame this benchmark-centered regime as Silicon Bureaucracy and AI Test-Oriented Education, and argue that it rests on a fragile assumption: that benchmark scores directly reflect genuine generalization. In practice, however, such scores may conflate exam-oriented competence with principled capability, especially when contamination and semantic leakage are difficult to exclude from modern training pipelines. We therefore propose an audit framework for analyzing contamination sensitivity and score confidence in LLM benchmarks. Using a router-worker setup, we compare a clean-control condition with noisy conditions in which benchmark problems are systematically deleted, rewritten, and perturbed before being passed downstream. For a genuinely clean benchmark, noisy conditions should not consistently outperform the clean-control baseline. Yet across multiple models, we find widespread but heterogeneous above-baseline gains under noisy conditions, indicating that benchmark-related cues may be reassembled and can reactivate contamination-related memory. These results suggest that similar benchmark scores may carry substantially different levels of confidence. Rather than rejecting benchmarks altogether, we argue that benchmark-based evaluation should be supplemented with explicit audits of contamination sensitivity and score confidence.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.