Looped inference extends task horizons but doesn't boost undetected sabotage, study finds
heghbalz · x · 2026-11-21
- Safety researchers (Alex Robey et al.) tested looping a model through multiple passes without written notes
- Loops let the model accomplish more: No-CoT 50% time horizon grew from 16s (1 pass) to 26s (2), 37s (3), 42s (5), 40s (10 passes), approaching Inkling 2 Mini's 41s at 3 passes
- Chain-of-thoughts got shorter but remained legible
- Key safety finding: models' ability to carry out sabotage-style tasks undetected did not meaningfully improve with added loops
More from Safety
- 'Imagine if Washington handed the internet to Gates in 1998': an AI regulation warning — Dan_Jeffries1 · 2026-11-22
- Can Zscaler Become the Default Control Plane for Enterprise AI Agent Access? — thedealdirector · 2026-11-22
- Solo Developer Builds Kavdren, a Verification Layer for AI Agent Responses — Few-Ad-1358 · 2026-11-22
- Anthropic's $1.5B book settlement unpaid; call grows to raise it to $20K per book — TuhinChakr · 2026-11-22
- AI Safety Researcher: Positive AI Futures Require Transparent US-China Collaboration — LuizaJarovsky · 2026-11-22
- NAS Report on Scientific Fraud Itself Contains Hallucinated Citations — KordingLab · 2026-11-22