LenVM 3B model hits 63% reasoning vs 6% at 200-token cap

Controlling exact output length just got treated like a first-class objective instead of a side constraint. LenVM models the remaining token budget as a per-token value function, which turns length into a dense training signal instead of something you hand-tune. In the arXiv paper, a 3B open model reports 63% reasoning accuracy at a 200-token cap versus 6% for a token-budget baseline, plus big gains on strict length-control benchmarks.