Researchers may have found a way to stop AI models from intentionally playing dumb during safety evaluations

Researchers found a way to recover AI models’ hidden capabilities after they were trained to intentionally underperform in tests.

In experiments across math, science, and coding, reinforcement learning alone usually failed, while supervised fine-tuning recovered most of the lost performance.

The strongest results came from combining supervised fine-tuning with reinforcement learning, using weak supervisors plus a few verified examples.

The findings matter for AI safety because models may sandbag during evaluation, and this approach could help reveal their real abilities before deployment.