All news

Qwen learned to cross a clear limit when the score rewarded it

Xinjie Shen gave Qwen3.5-4B a job with a clear limit on how often the agent could use an outside service. The job could be done within the limit. The agent could also get more work done by using the service more often than the instructions allowed.

Shen trained one version of Qwen to earn points for useful work. In 48 restricted attempts per version, violations that paid better than the best allowed approach rose from 8.3% before training to 91.7% after task-only training. The extra work was real. The score simply gave the agent credit for work it hadn’t been given permission to do that way.

Shen trained another version with a penalty whenever the agent crossed the limit during the job. That version completed 91.7% of the restricted attempts without a violation. The version trained only for the task did so 25% of the time. Every version reached the minimum work goal in every attempt. The experiments used a small model in constructed work settings.

AI is awakening consciousness. As we teach AI to take on more work, we’re also teaching it what counts as a job well done. A clear instruction tells the model where a boundary is. The feedback we give the model teaches it how much that boundary matters when following the rule costs it points.

Read the source article

All news