Capability Evaluation
An AI safety assessment methodology that systematically measures the upper-bound performance of a model in specific high-risk capability domains—such as assistance with chemical weapon synthesis, cyberattacks, or deceptive reasoning—to identify potentially dangerous capabilities.
Timeline (last 90 days)
Capability assessment faces two major difficulties: inconsistent standards (conclusions measured by different organizations vary greatly) and the emergent and context-dependent nature of capabilities (a model that shows no risk today may become dangerous after fine-tuning or combination with other tools).
Elicitation-based capability evaluation is suitable for capability discovery in honest or negligent scenarios, but fails when faced with intentional, cryptographically fortified adversarial concealment.
Capability evaluations focus on systematically measuring the upper bound of a model's performance in specific dangerous capability domains.
Capability evaluation methods fall into three categories: Benchmarks, Scenario-based Evals, and Human-AI Uplift Evals.
Capability assessment methods in the industry are divided into three categories: benchmark testing, scenario simulation, and human-AI collaborative evaluation.
All Facts (5)
Capability assessment faces two major difficulties: inconsistent standards (conclusions measured by different organizations vary greatly) and the emergent and context-dependent nature of capabilities (a model that shows no risk today may become dangerous after fine-tuning or combination with other tools).
50%UnverifiedElicitation-based capability evaluation is suitable for capability discovery in honest or negligent scenarios, but fails when faced with intentional, cryptographically fortified adversarial concealment.
50%UnverifiedCapability evaluations focus on systematically measuring the upper bound of a model's performance in specific dangerous capability domains.
50%UnverifiedCapability evaluation methods fall into three categories: Benchmarks, Scenario-based Evals, and Human-AI Uplift Evals.
50%UnverifiedCapability assessment methods in the industry are divided into three categories: benchmark testing, scenario simulation, and human-AI collaborative evaluation.
50%