Researchers propose using general instructions as privileges in On-Policy Context Distillation (OPCD) to improve out-of-distribution (OOD) performance and maintain in-domain performance. Experimental results on various datasets and models show that matched instruction privileges outperform gold privileges in OOD accuracy.

MetaCtrl is a lightweight controller that adaptively regulates large language models (LLMs) to improve their reasoning accuracy while reducing inference-time generation. It observes the evolving reasoning trace, decides whether to continue, simplify, or conclude reasoning, and is trained using reinforcement learning. MetaCtrl improves the accuracy of LLMs on various benchmarks, including mathematics, science, and code, and can transfer to unseen reasoners without further training.