Controllable Reasoning Models Are Private Thinkers

Researchers developed a method to train AI reasoning models to follow privacy instructions in their internal reasoning traces, not just final answers. The approach uses separate LoRA adapters and achieves up to 51.9% improvement on privacy benchmarks, though with some trade-offs in task performance.

• →New training method enables AI models to follow privacy constraints during internal reasoning processes, not just in final outputs.
• →The approach uses separate LoRA adapters to decouple reasoning and answer generation for better control.
• →Testing across models ranging from 1.7B to 14B parameters showed up to 51.9 percentage point improvements on privacy benchmarks.
• →Method achieved up to 20.9 point gains in instruction-following performance across multiple benchmarks.
• →Privacy improvements come with trade-offs in task utility, highlighting the balance between reasoning performance and privacy preservation.