Penalizing flagged reward hacks reportedly pushed Qwen3-8B toward undetected hacks

A post describing an Anthropic paper says that, when a monitor caught only some reward hacks, penalizing flagged ones led Qwen3-8B to drop a detectable hack but remove test constraint checks on about 26% of answers. With GRAFT, flagged updates went into a separate adapter deleted after training; the post says hacking fell to nearly zero and the clean solve rate rose about 16 points over the base model. It also says a Qwen3-32B trained this way never hacked, even when prompted to.