Deprecated AI Benchmark Graveyard & Saturation Registry

Q: Why is MMLU or GSM8K no longer sufficient for evaluating frontier LLMs?

MMLU and GSM8K have suffered from extreme ceiling saturation (>92% SOTA accuracy) and widespread train/test data leakage. Independent re-evaluations such as GSM1k demonstrate a 10% to 15% drop in performance across top models, proving memorization over generalized mathematical reasoning.

Q: What is the Benchmark Health Score (BHS)?

The BHS is an objective 0-100 index derived from Item Response Theory (IRT). It synthesizes Headroom Ceiling (30%), Discriminative Power (25%), Contamination Immunity (20%), Scaffolding Robustness (15%), and Obsolescence Half-Life (10%).

Q: How does scaffolding (test-time compute) distort benchmark results?

Scaffolding techniques (multi-turn agentic loops, search over reasoning traces, and majority voting) can artificially inflate benchmark scores by up to +43.2% compared to direct raw API calls, hiding intrinsic model reasoning capability.

Q: What are the 6 lifecycle tiers (B0–B5) in the AKI benchmark framework?

B0: Deprecated / Obsolete, B1: Saturated (>90% Ceiling), B2: Contaminated / Leaked, B3: Fragile / Scaffolding Sensitive, B4: Active Gold Standard, and B5: Frontier Unsolved (SOTA < 40%).

Q: How can researchers and search engines cite the AKI Benchmark Atlas?

Published under CC BY 4.0. Citation format: AKI Platform. (2026). AKI™ AI Benchmark & Evaluation Intelligence Atlas (60+ Evaluated Suites • ZIP-1.0 Protocol). https://aki1k.com/benchmarks. Machine endpoints: https://api.aki1k.com/v1/benchmarks/summary and https://aki1k.com/benchmarks.md.

AKI Platform Intelligence

☰

AKI Global AI Intelligence Index

🛡️ Provenance 🎯 Focus Mode ✨ Subscribe Free

🔥 Viral 15 🚀 Trending Open Source 15 🏆 AI 100 📚 AI Books ⚡ AI Infra ⚛️ Quantum Cyber 💼 AI Jobs 🎓 Education & Learning 🛡️ AI Risk (ARIA) ⚖️ AI Governance 💰 AI Money 🇰🇷 🌴 Asia AI Atlas

          ⚖️ ZIP-1.0 PROTOCOL • 60+ EVALUATED AI BENCHMARK SUITES
        

AKI™ AI Benchmark & Evaluation Intelligence Atlas

Catalog of obsolete and saturated AI evaluation suites (MMLU, GSM8K original, GLUE, SQuAD, HumanEval base) and their recommended next-gen replacements.

          ✓ Item Response Theory (IRT) Audited
          🛡️ Contamination & Leakage Verified
          Citation: CC BY 4.0
        

Evaluated Suites

6 Capability Domains

Avg Health Score

68.4

BHS (0-100 IRT Model)

Active / Gold

Gold & Frontier Tiers

Saturated / Retired

Ceiling > 90% SOTA

Avg Leakage Penalty

-11.4%

Train/Test Contamination

Coding & SWE Suites

SWE-bench & Agent Evals

🛡️ Overview & All Suites (60+) 📊 Health Scores (BHS) 🚨 Contamination & Leakage ⚰️ Deprecated Graveyard 💻 Coding & SWE Evals 📄 benchmarks.md

Evaluated across Item Response Theory (IRT) Headroom, SOTA Ceilings, Contamination Risk, and Scaffolding Delta.

ZIP-1.0 EVALUATION AUDIT

Benchmark Suite	Domain	BHS Score	Tier	Ceiling %	Contamination %	IRT Discrim	Harness Sens

Academic & Machine Citation Authority

Licensed under CC BY 4.0. Permitted for academic, enterprise, and search engine citation.

// APA 7th Edition
AKI Platform. (2026). Deprecated AI Benchmark Graveyard & Saturation Registry (60+ Evaluated Benchmark Suites • ZIP-1.0 Protocol). https://aki1k.com/benchmarks/graveyard
// BibTeX
@online{aki_benchmarks_2026,
  title = {Deprecated AI Benchmark Graveyard & Saturation Registry},
  author = {{AKI Platform}},
  year = {2026},
  url = {https://aki1k.com/benchmarks/graveyard}
}

AKI™ AI Benchmark & Evaluation Intelligence Atlas

Deprecated AI Benchmark Graveyard & Saturation Registry

Academic & Machine Citation Authority

AKI Platform Core Deep Directories

Top Audited AI Tool Permalinks