Eliciting Latent Knowledge
The AI-safety problem of reliably extracting human-relevant facts that a capable model internally represents, even when the model's ordinary outputs or available observations can be misleading and direct ground-truth supervision is unavailable; in the original framing, this requires mapping between the model's world model and a human's concepts.
Metadata
- aliases
- en
- ELK
- concept_id
- eliciting_latent_knowledge
- preferred
- en
- Eliciting Latent Knowledge
- relations
- status
- active