LLM-guided headline rewriting for clickability enhancement without clickbait

Optimizing news headlines for reader engagement is often conflated with clickbait, producing exaggerated or misleading phrasing that erodes editorial trust. We reframe clickbait not as a separate category but as the disproportionate amplification of otherwise legitimate engagement cues, and cast headline rewriting as a controllable generation problem in which chosen engagement attributes are strengthened under explicit faithfulness constraints. We steer a large language model at inference time with two lightweight guide models: a clickbait scorer providing negative guidance and an engagement-attribute model providing positive guidance, combined during decoding so a chosen engagement attribute can be selectively strengthened while clickbait is restrained. Both guides are trained on neutral news headlines and their synthetically amplified variants; the clickbait guide is externally validated on human-authored corpora, and the engagement guide’s attribute scores correlate with human clickbait. On held-out headlines, an ablation shows the two guides exert independent, opposite effects, and benchmarked against the decoding-time control methods DExperts and GeDi the dual guide adds an independent clickbait brake, a dedicated lever that lowers induced clickbait at a chosen engagement level and that single-axis methods lack. Independent automatic metrics and a three-annotator human study confirm higher faithfulness and lower induced clickbait than these baselines. The framework supports responsible headline optimization in journalism. The setting we report keeps rewrites faithful to the source and low in clickbait rather than pushing engagement as high as possible; a lighter brake is available when more engagement is wanted.