Token-Level Length Control: 3B Model Beats GPT 5.4 and Claude

token级，精准控制生成长度：3B模型击败GPT 5.4、Claude

UC Santa Barbara and Apple researchers introduced LenVM, which models remaining generation length as a token-level value function; Qwen2.5-3B with a 1.5B LenVM scored 62.6 on LIFEBench length control, above GPT-5.4 at 37.4 and Claude-Opus-4-6 at 35.5.

Why it matters: HKR-H/K/R all pass: the headline has a sharp small-model-vs-frontier hook, and the post gives LenVM's mechanism plus 62.6/37.4 benchmark numbers. The topic is narrow research, not a model or major product release, so it fits the 78-84 band.