---
aliases:
  en:
  - inner alignment problem
concept_id: inner_alignment
preferred:
  en: inner alignment
  ja: インナーアラインメント
relations:
  relations:
    broader:
    - ai_alignment
    prerequisite:
    - ai_alignment
    - base_objective
    - mesa_objective
    - mesa_optimizer
    related:
    - objective_robustness
    - outer_alignment
status: active
---

# inner alignment

The problem or property of ensuring that a mesa-optimizer's mesa-objective is robustly aligned with the base objective used by the base optimizer, rather than merely producing high-scoring behavior on the training distribution. In objective-focused terminology this is the mesa-optimizer special case of objective robustness; the exact scope of inner alignment varies across authors and settings.

## broader

- [AI alignment](../terminology/ai_alignment.txt)

## prerequisite

- [AI alignment](../terminology/ai_alignment.txt)
- [base objective](../terminology/base_objective.txt)
- [mesa-objective](../terminology/mesa_objective.txt)
- [mesa-optimizer](../terminology/mesa_optimizer.txt)

## related

- [objective robustness](../terminology/objective_robustness.txt)
- [outer alignment](../terminology/outer_alignment.txt)

## Documents: about

- [ベース目的関数とメサ目的関数は一致するとは限らない](../documents/hubinger_et_al.base-and-mesa-objectives-can-diverge.txt)
- [観測された従順さだけでは、内部の目的や欺瞞の有無までは分からない](../documents/yudkowsky.observed-compliance-does-not-identify-hidden-objectives.txt)
- [外部の目的関数は内部の目標を決定しない](../documents/yudkowsky.outer-objective-does-not-determine-inner-goal.txt)

## Documents: mentions

- [能力はアラインメントより遠くまで汎化しうる](../documents/yudkowsky.capabilities-generalize-beyond-alignment.txt)
- [能力と目的は別の軸である](../documents/yudkowsky.capability-motive-separation.txt)

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