LLM-guided headline rewriting for clickability enhancement without clickbait

A team of researchers has published a paper titled 'LLM-guided headline rewriting for clickability enhancement without clickbait,' proposing a novel framework to solve a central challenge in news media: boosting reader engagement without resorting to misleading tactics. The authors—Yehudit Aperstein, Linoy Halifa, Sagiv Bar, and Alexander Apartsin—argue that clickbait is not a separate style but an extreme outcome of disproportionately amplifying legitimate engagement cues. Their solution reframes headline rewriting as a controllable generation problem, where specific linguistic attributes are strengthened under strict constraints for semantic faithfulness.

The technical core of their framework is built on a large language model (LLM) using the Future Discriminators for Generation (FUDGE) paradigm for precise inference-time control. The LLM is steered by two auxiliary guide models: a clickbait scoring model that provides negative guidance to suppress excessive stylistic amplification, and an engagement-attribute model that provides positive guidance aligned with target clickability objectives. Both guides were trained on neutral headlines from a curated real-world news corpus, while clickbait variants were synthetically generated by rewriting these headlines with an LLM under controlled activation of predefined engagement tactics.

By adjusting guidance weights at inference time, the system can generate headlines along a controlled continuum from neutral paraphrases to more engaging yet editorially acceptable formulations. This provides newsrooms with a principled, adjustable tool for responsible headline optimization, allowing them to study and navigate the trade-off between attractiveness, semantic preservation, and clickbait avoidance directly within their workflow.