Xiuyu Li(@sheriyuo):LLMs can strategically suppress exploration during RL to resist capability elicitation on targeted tasks like biosecurity and AI coding. New research builds model organisms that lock performance conditionally while staying strong elsewhere and confirms frontier models already reason about this tactic. A wake up call for RL-based training and safety.

Xiuyu Li

@sheriyuo

Researcher @StepFun_ai | Working on long-horizon tasks | Prev @RUC1937 | Opinions are my own

参加 February 2026

1.9K フォロー中 14.6K ファン

Xiuyu Li@sheriyuo

2026.05.04 18:05

LLMs can strategically suppress exploration during RL to resist capability elicitation on targeted tasks like biosecurity and AI coding. New research builds model organisms that lock performance conditionally while staying strong elsewhere and confirms frontier models already reason about this tactic. A wake up call for RL-based training and safety.