#held-out benchmarks

Tagged held-out benchmarks · show all articles

A sealed task cartridge caught mid-slide into a dock on a rotating carousel inside a bright vault of shelved cartridges, its foil seal intact and its latch not yet closed

How Benchmark Contamination Actually Works in Agentic Evaluation

A single-turn quiz question can be memorized whole. An agentic benchmark is an entire environment — a repository, a test suite, a live shell — and each of those can leak, get gamed, or get recognized as a test in ways a static question bank cannot.

By Brecht Corbeel · August 12, 2026 · 18 min read

Browse by Category

Surprise me →

All AI Hardware & Semiconductors 100 AI Agents & Systems 33 Foundation Models 27 Model Evaluation 15 AI Industry 9 Edge AI & Electronics 7 AI Safety 7 AI Research 7 Open Models 6 AI Infrastructure 5 Industry Analysis 2 Datacenters & Infrastructure 2 Writing & Information 1 Structural Biology 1 Small & Edge Models 1 Semiconductor Markets 1 Security & Technology 1 Security 1 Scientific Revolution 1 Robotics & Embodied AI 1 Quantum Information 1 Planetary Science 1 Physics 1 Particle Physics 1 Origins & Astrobiology 1 Open-Weight Models 1 Neuroscience 1 Metrology 1 Mathematics & Computation 1 Materials Science 1 Labour & Technology 1 Interpretability 1 Institutions & State Capacity 1 Institutions & Economy 1 Inference Economics 1 Industrial Revolutions 1 Humanity & Technology 1 History of Technology 1 History of Computing 1 Governance & Policy 1 Evolution & Ecology 1 Ethics of Technology 1 Energy & Civilization 1 Earth & Climate 1 Demography & Urbanization 1 Data & Training 1 Cumulative Culture 1 Cosmology 1 Condensed Matter 1 Chemistry & Catalysis 1 Alignment & Safety 1 Agriculture & Cities 1 Agent Evaluation 1 AI Security 1 AI & The Web 1