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

Researchers introduce LenVM, a token-level value model for remaining generation length.

It uses a constant negative reward per token to create dense, annotation-free length signals at inference time.

LenVM improves exact length matching, supports budget-based tradeoffs, and works across LLMs and VLMs.

The approach also helps explain how individual tokens affect generation length.