Masked Distillation: Internalizing Chain-of-Thought to Cut LLM Inference Costs

rao2z · x · 2026-07-31

Large Reasoning Models (LRMs) generate long, explicit chains of intermediate steps before final answers, heavily increasing latency, memory usage, and serving costs. Researchers from ASU and others introduced the masked distillation framework to internalize these intermediate computations into model parameters.

The framework operates in two modes: self-distillation (the same model acts as a teacher in thinking mode and a student in non-thinking mode) and dual-model distillation (a larger reasoning model supervises a smaller non-thinking model). This approach offers a train-inference tradeoff: by over-training the base model on specific problem distributions, the intermediate tokens produced during inference can be significantly reduced.

Related event: Masked Distillation Framework Aims to Reduce LLM Inference Costs(2 posts)→

Original post →

More from Research

Research channel →