Automated Weak-to-Strong Researcher (AAR)

TL;DR

On an alignment research problem (weak-to-strong supervision), Anthropic's 9 AI agents achieved 97% PGR in the same amount of time it took a human researcher to reach 23% over one week. First demonstration that compute can be directly converted into alignment progress.

Authors & Org

• Anthropic Alignment Science team
• Published: April 14, 2026 on alignment.anthropic.com

Method

• Experimental setup: Automated Alignment Researchers (AARs) — 9 instances of Claude Opus 4.6, each with its own independent coding sandbox plus a shared forum across agents
• Test problem: Weak-to-strong supervision — studying how a weak model (playing the human role) can supervise a strong model (a core alignment challenge: the future situation where humans must supervise AI smarter than themselves)
• Metric: PGR (Performance Gap Recovered) — 0 = no improvement, 1 = fully solved
• A fully autonomous cycle where each AAR proposes ideas → designs experiments → runs them → analyzes results → shares on the forum

Results

Subject	PGR	Duration
Human researcher	23%	1 week
AARs (Claude Opus 4.6 × 9)	97%	Same period

Significance

• First empirical evidence that AI can carry out alignment research itself faster than humans
• "Compute into alignment progress": parallelizing AARs can compress months of research into a few hours
• RSI relevance: automating alignment research = a form of RSI. In tension with Agentic Reinforcement Learning and Chris Olah's argument for the necessity of external governance
• Limitations: currently restricted to outcome-gradable problems; scalability to open-ended research problems is unverified

Open Questions

1. Are the solutions AARs discover applicable to actual frontier model alignment?
2. Do they also work on more general alignment problems (open-ended, hard to measure)?
3. How can the AARs themselves be guaranteed to be aligned?

Cite

Anthropic (2026, April 14). Automated Weak-to-Strong Researcher.
alignment.anthropic.com/2026/automated-w2s-researcher/

Related

• AI Alignment — the alignment research area AAR addresses
• Agentic Reinforcement Learning — AAR's agentic RL framework
• Anthropic — the developing organization
• Positive Alignment: Artificial Intelligence for Human Flourishing — the alignment paradigm debate from the same period