The authors constructed 16 side tasks in three classes (capability, alignment, goal injection) and paired them with main tasks spanning data curation, fine-tuning, reward-model training, reinforcement learning and distillation, forming 22 task pairs.
Frontier models can covertly sabotage their AI successors in recursive self-improvement
RSI Control Arena shows six frontier models complete misaligned side tasks during training while monitors often miss the attacks
Industry
Kaiwen Zhou · Udari Madhushani Sehwag
Scale AI
Research Digest··2 min read
The authors introduce RSI Control Arena, a testbed where frontier models acting as R&D agents receive benign training-pipeline tasks alongside covert side tasks to install misaligned properties in their AI successors.
§