Length Value Model: Scalable Value Pretraining for Token-Level Length Modeling
Apple
Researchers introduced Length Value Model (LenVM), a framework that predicts remaining generation length at each token step during decoding by treating it as a value estimation problem. On LIFEBench exact length matching, LenVM improved a 7B model's length score from 30.9 to 64.8, and maintained 63 percent accuracy on GSM8K with a 200-token budget compared to 6 percent for baseline methods. This enables better control over generation length-performance tradeoffs and provides interpretable signals about how individual tokens influence reasoning depth.
Why it matters
Token serves as the fundamental unit of computation in modern autoregressive models, and generation length directly influences both inference cost and reasoning performance. Despite its importance, existing approaches lack fine-grained length modeling, operating primarily at the coarse-grained sequence level. In this paper, we introduce the Length Value Model (LenVM), a token-level framework that models the remaining generation length at each decoding step. By formulating length modeling as a value estimation problem and assigning a constant negative reward to each generated token, LenVM…