kapynResearch

Length Value Model: Scalable Value Pretraining for Token-Level Length Modeling

Length Value Model is a token-level framework that predicts remaining generation length during decoding. The method frames length modeling as a value estimation problem by applying constant negative rewards at each decoding step. This approach gives developers fine-grained control over inference costs and reasoning performance compared to traditional sequence-level methods.

Apple ML Research·Jul 20, 2026

Opening Kapyn…