NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·May 26·5 min read
Retrying vs Resampling in AI Control
TL;DR
AI coding tools like Claude Code and Codex use a method called retrying to handle risky actions. This study introduces resampling, which involves drawing multiple samples from the same context, showing it can enhance safety without leaking exploitable information.
✦ Why It Matters
Engineers can adopt resampling techniques to enhance AI safety without sacrificing performance.
Key Takeaways
- Evaluate the use of resampling over retrying in AI control systems.
- Implement maximum suspicion score auditing for better safety outcomes.
- Consider drawing multiple samples to mitigate risks from untrusted models.
- Reassess the effectiveness of executing the least suspicious sample in your protocols.
How It Works
Resampling involves generating multiple outputs from the same input context, allowing for a more robust evaluation of potential risks. By auditing the maximum suspicion score from these outputs, the system can better identify and mitigate threats, as opposed to relying on the minimum score, which may overlook critical vulnerabilities.