Apple Proposes Token-Level Length Value Model to Optimize Inference Cost

Apple ML Research · rss · 2026-07-20

Apple's machine learning team introduced the Length Value Model (LenVM), a novel framework designed to address how generation length impacts both inference cost and reasoning performance in autoregressive models.

While existing approaches primarily operate at a coarse-grained sequence level, LenVM introduces fine-grained, token-level length modeling. By formulating length modeling as a value estimation problem, the framework predicts the remaining generation length at each decoding step and assigns a constant negative reward to each generated token, enabling more scalable and precise pretraining.

Original post →

More from Research

Research channel →