Simple features beat most AI control monitors
Original title: Simple Features Beat Most AI Control Monitors
Simple features separate honest from backdoored code more effectively than most monitors, including Claude Sonnet 5.
On a standard evaluation dataset, simple features like code style and problem characteristics outperform most monitors without executing code. Human-written honest solutions and backdoors from older models allow style-based separation. Swapping classes with Claude Sonnet 5 code boosts contrast between authors to 0.95 from 0.79. One in seven backdoors fails validity checks.
Unconfirmed: Whether published monitors rely on these shortcuts is unconfirmed.
- Simple features achieve 0.95 separation vs 0.79 original
- One in seven backdoors invalid
- Sonnet 5 monitor catches 77% older backdoors but 37% own
- 78% vs 61% after verification
- Dataset rebuilt with checks released
From the original article
1 min read at arxiv.orgAI control relies on monitors to catch a capable model that sabotages its work. These monitors are evaluated on their ability to differentiate honest from backdoored code. We show that on a standard evaluation dataset, simple features beat most monitors without ever running the code, outranking 20 of 24 open-weight models prompted as monitors.
By Shubhrangshu Debsarkar · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.