Safety
Latent Correctness Direction (LCD)
2025ResearchUpdated: 22 September 2026Published
Key
innovation
Identifying a linear 'correctness' direction encoded in a model's activations.
Category
Safety
Abstraction level
Pattern
Operation level
ModelInference
Use cases
Hallucination detectionConfidence estimationActivation steering
How it works
Probing finds a direction in activations separating correct from incorrect answers; it can be used for detection or steering.
Problem solved
Models often 'know' internally that they are wrong, yet this information is not directly available in the output.