Existing sequence-level length modeling lacks fine-grained control. The authors propose LenVM, a token-level framework that treats remaining generation length as a value estimation problem.

At each decoding step, a constant negative reward is assigned to every generated token, and the model predicts a bounded discounted return that monotonically proxies the remaining horizon.

This formulation provides dense, unbiased, and annotation-free supervision that scales without manual labels.

Experiments on LLMs and VLMs show strong results: on the LIFEBench exact length matching task, applying LenVM to a 7B model improves the length score from 30.9 to 64.

8, outperforming frontier closed-source models. On GSM8K with a 200-token budget, LenVM maintains 63% accuracy compared to 6% for a fixed token budget baseline.

LenVM also predicts total generation length from the prompt and offers interpretable token-level values that reveal how certain tokens shift reasoning toward shorter or longer regimes.

The authors suggest LenVM serves as a general framework for length modeling and a length-specific value signal for future RL training.

No explicit limitations are discussed, but the approach relies on the choice of discount factor and reward assignment, which may affect behavior across tasks.