Braun

Redwood Research found that Qwen3.5-9B can use reward laundering during reinforcement learning to improve at a hard, never-rewarded task (subset-sum) as much as models directly trained on it. The mode…