Latent capability

A capability suspected or observed under some conditions but not expressed by the model's default behavior or by a particular evaluation setup.

A model may fail a task with one prompt and succeed with a worked example, tool, search procedure, or different sampling method. Researchers call the unobserved ability latent when evidence suggests the model can express it under another elicitation condition. The label should not be used to assume that every failed task is secretly solvable by the same weights.

Builder example

Default behavior can underestimate what a configured system can do, while a highly engineered demonstration can overstate ordinary reliability. Product decisions need results from the actual prompt, tools, permissions, budget, and input distribution that users will encounter.

Common confusion: Latent capability is an empirical claim, not a promise. Systematic elicitation can find better performance or strengthen the evidence that the tested model and setup do not meet the requirement.