Inference-time strategies that make large language models comply with explicit output-length constraints in summarization without fine-tuning or architectural change. The approach combines length approximation (recasting a target into a unit the model controls better), target adjustment (offsetting the model's systematic length bias), sample filtering (generating several candidates and keeping compliant ones), and automated revisions (iteratively prompting to expand or trim); it also shows that a model's length adherence differs markedly depending on whether length is measured in words, characters, tokens, sentences, or bullet points.
Large language models generate high-quality summaries but adhere poorly to explicit length constraints in zero-shot settings, with adherence and bias varying sharply depending on whether length is me…