Neel Nanda: no alignment faking found in Gemini 4 Argon RL training audit of 10K rollouts
NeelNanda5 · x · 2026-10-17
Responding to replication questions, DeepMind's Neel Nanda reports that audits of Gemini 4 Argon RL training — all monitor-flagged transcripts plus a stratified random sample of 10,000 multi-step agentic rollouts (11,334 context spans) — found zero instances of strategic alignment faking or coherent beyond-episode drives (self-preservation, goal integrity, resource acquisition), and no prompted sabotage, though high rates of evaluation awareness were observed.
Related event: DeepMind finds no alignment faking in Gemini 4 Argon training audit(2 posts)→
More from Models
- Kaggle launches a 'pelican riding a bike' arena: GPT-6 Astra leads, Gemini 4 Argon close behind — MeganRisdal · 2026-10-17
- Dev Slams Corporate HR-Style AI: I Want a Model That Tells Me When I'm an Idiot — VraserX · 2026-10-17
- Perplexity's Decider V1.2 decision models land on Vercel AI Gateway — AravSrinivas · 2026-10-17
- Grok Bot update: 3x faster computer tasks, X Chat access, new connectors and plugin marketplace — XFreeze · 2026-10-17
- Replicating an interpretability paper with Opus 5.5: 20 of 22 results confirmed on Qwen3-32B — Sauers_ · 2026-10-17
- Chinese Frontier Models Go Quiet: Kimi K3 at 92 Days, K3.1 and GLM-5.5 Rumored Soon — haider1 · 2026-10-17