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

Introduced the Length Value Model (LenVM), a novel token-level framework for modeling remaining generation length. LenVM estimates length as a value prediction problem with a constant negative reward per token, yielding a discounted return proxy for the generation horizon. This approach provides annotation-free, dense, unbiased, and scalable supervision. Demonstrated significant improvements in exact length matching on LIFEBench (from 30.9 to 64.8 for a 7B model) and enabled effective performance-efficiency trade-offs on GSM8K (maintaining 63% accuracy at a 200-token budget vs. 6% for a baseline). Provided interpretable token-level values to understand generation dynamics and their influence on length.