// ⛨ Safety, Evaluation & Security

Weak-to-Strong Generalization

Supervising strong models with weak-model labels as an analogy for superalignment.

弱到强泛化 2023 Advanced core

IN PLAIN WORDS

GPT-2-level labels supervise a GPT-4-level student; the student beats its teacher but not its full potential, and an auxiliary confidence loss narrows the gap.

Used in: Scalable oversight research

TIMELINE · prerequisites → this → what builds on it

2015 Knowledge Dis…2022 Reinforcement…2023 Weak-to-Strong Ge…

01Original paper

Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision

Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, et al. · 2023

arXiv ↗ ✓ Verified · arXiv API

02Where it sits

Learn the left column first; the right column builds on this one.

Learn first

Weak-to-Strong Generalization

Then you can learn

Nothing in the atlas builds on it yet

Related:AI Safety via Debate

Left: prerequisites · right: builds on this · dashed: related. Full map →

03Learning resources

Official docs, explainers, courses and reference code. All links checked.

Explainers

OpenAI：Weak-to-strong generalization ↗EN

04Where the course teaches it

U93 Frontier Risk, Scalable Oversight & Governance Weak-to-strong generalisation, debate, Constitutional AI, risk frameworks and regulation (EU AI Act, NIST AI RMF). 6 h4 algorithms3 papers

For AI / data

Verification: verified, checked 2026-10-21.
JSON: /api/learn/algorithms.json (id: weak-to-strong)