LLM Inference Efficiency Gains May Just Be Overtraining Base Models

rao2z · x · 2026-07-31

A researcher points out that many current methods aiming to reduce the length of intermediate tokens during LLM inference might essentially just be "overtraining" the base model on specific distributions.

The thread reviews a recent paper by @durgeshkalwar et al. that tests this hypothesis using Masked Distillation. By masking parts of the teacher model's trace during distillation, the student model is forced to "internalize" intermediate token generation. Experiments show full internalization works well on GSM8K, significantly reducing token length. However, in the Countdown task, full internalization degrades performance, while partial internalization retains performance with fewer tokens. This offers a new train-inference tradeoff perspective on LLM efficiency.

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

Original post →

More from Research

Research channel →