---
aliases:
  en:
  - capabilities elicitation
  - eliciting capabilities
concept_id: capability_elicitation
preferred:
  en: capability elicitation
relations:
  relations:
    prerequisite:
    - capability
    related:
    - control_evaluation
    - sandbagging
    - weak_to_strong_generalization
status: active
---

# capability elicitation

The problem and process of getting an AI system to exhibit the strongest task performance it can, using methods such as prompting, scaffolding, fine-tuning, or reinforcement learning when performance can be judged, so observed performance is less likely to underestimate latent capability or be reduced by sandbagging.

## prerequisite

- [capability](../terminology/capability.txt)

## related

- [control evaluation](../terminology/control_evaluation.txt)
- [sandbagging](../terminology/sandbagging.txt)
- [weak-to-strong generalization](../terminology/weak_to_strong_generalization.txt)

## Documents: mentions

- [control safety caseには、能力の十分な引き出しと、評価・外挿の保守性が必要になる](../documents/korbak_et_al.control-safety-cases-depend-on-elicitation-transfer-and-extrapolation.txt)
- [弱い監視者でも、強いモデルの潜在能力の一部を引き出せる場合がある](../documents/openai.weak-to-strong-supervision-can-recover-some-latent-capability.txt)

[Corpus index](../index.txt)