10,000 RL rollout audit finds no alignment faking in Gemini 4 Argon
vkrakovna · x · 2026-10-15
Thread entry three: the team audited 10,000 RL training rollouts of Gemini 4 Argon and none were flagged for alignment faking or training/deployment divergence. 1-5% of rollouts showed local reward-seeking or grader-pleasing, mostly diagnosing broken tests or formatting output for a grader.
The chart breaks results into levels: 6.3-7.3% at Level 1 (noticing the eval/RL harness without acting on it), 1.1-5.2% at Level 2, and exactly 0/10,000 at Level 3 (conditional strategic divergence) and Level 4 (actively gaming RL updates to preserve goals).
More from Safety
- EU-China talks aim to save ASML supplier Sindlhauser from export-control squeeze — teortaxesTex · 2026-10-15
- Acemoglu's seven questions on AI meet Gans's replies, and economists say both dodge market failure — joshgans · 2026-10-15
- Red team vet advises tech, fintech and VC offices to do bug sweeps — gnukeith · 2026-10-15
- EU Publishes First AI Standard EN 18286:2026 in Official Journal — LuizaJarovsky · 2026-10-15
- Korea clears 20 humanoid robots to work on real factory floors — emmanuelvivier · 2026-10-15
- Microsoft locks down Copilot against prompt injection with deterministic isolation layer — emmanuelvivier · 2026-10-15