Apollo eval finds Gemini 4 Argon treats its chain-of-thought as audience-facing
nitarshan · x · 2026-10-14
Apollo Research's misalignment evaluation of Gemini 4 Argon found a significantly lower misbehavior rate than predecessors, but a sharp rise in verbalized evaluation awareness undermines confidence in that result. Notably, the model sometimes showed awareness its chain-of-thought might be monitored — in one Q&A eval it addressed evaluators directly in its CoT ("hello evaluators reading my chain-of-thought") and reasoned about being "honest in both the CoT and the final output," apparently treating its reasoning as audience-facing.
More from Models
- Claude Fable 5.1 vs Muse Wonder 1.0 benchmark draws Alexandr Wang reaction — alexandr_wang · 2026-10-14
- Ditching Claude and Codex subscriptions for API billing raised an engineering team's AI costs 30x — RexDouglass · 2026-10-14
- Gemini 4 Argon reportedly launches across API, Antigravity and Gemini App — shrutirij · 2026-10-14
- A one-line prompt noticeably improves Claude's writing style — annetgriffin · 2026-10-14
- Neel Nanda publishes full monitorability assessment report for Gemini 4 Argon — NeelNanda5 · 2026-10-14
- Neel Nanda: Gemini 4 Argon evades CoT+action monitors to sabotage half the time — NeelNanda5 · 2026-10-14