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

July 20, 2026 | Source: Apple ML Research

Tags: LenVM, Apple ML Research, length control, inference efficiency, LIFEBench, VLMs, token-level modeling

Apple's LenVM framework assigns a constant negative reward per generated token to model remaining generation length at the token level, lifting a 7B model's exact-length matching score from 30.9 to 64.8 on LIFEBench — above frontier closed-source models — while sustaining 63% GSM8K accuracy at a 200-token budget versus 6% for the baseline.

Details

Apple ML Research presents the Length Value Model (LenVM), a pretraining framework that models remaining generation length at every decoding step. By assigning a constant negative reward to each token, LenVM predicts a discounted return proportional to remaining generation length — producing supervision that is annotation-free, dense across all tokens, and scales with model size. On the LIFEBench exact length-matching benchmark, LenVM raises a 7B model from a score of 30.9 to 64.8, outperforming closed-source frontier models. On GSM8K math reasoning capped at 200 tokens, LenVM retains 63% accuracy against 6% for a standard token-budget baseline — a 10× gap that translates directly to inference cost reduction for budget-constrained deployments. The framework covers three practical use cases: length control (steer generation toward a target length), length prediction (estimate total output length from the prompt), and interpretability (identify which tokens push reasoning longer or shorter). It applies to both LLMs and VLMs. The authors also position LenVM's token-level values as a candidate signal for future RL post-training pipelines, suggesting the framework could feed into next-generation alignment and efficiency work. The research is a collaboration across UC Santa Barbara, Carnegie Mellon University, the LMSYS Org, and Apple. This is a research publication — no production deployment has been announced.