LLM-guided headline rewriting for clickability enhancement without clickbait

1. Problem Statement

The paper addresses the structural tension in digital journalism between reader engagement (attracting clicks) and editorial integrity (maintaining factual accuracy).

• The Core Issue: Optimizing headlines for engagement often leads to "clickbait"—exaggerated, misleading, or sensationalist phrasing that erodes trust.
• Limitations of Current Approaches: Existing research treats clickbait as a binary classification problem (clickbait vs. non-clickbait) or focuses solely on detection/filtering. This approach fails to recognize that many engagement mechanisms (e.g., curiosity gaps, emotional framing) are legitimate rhetorical tools in journalism. They only become problematic when amplified disproportionately.
• The Gap: There is a lack of frameworks that can continuously control the degree of engagement to enhance clickability while strictly preventing the slide into clickbait, without requiring full retraining of Large Language Models (LLMs).

2. Methodology

The authors propose a controllable text generation framework based on the Future Discriminators for Generation (FUDGE) paradigm. The system rewrites neutral headlines into engaging ones by steering an LLM at inference time using two auxiliary models.

A. Conceptual Framework: The Engagement-Attribute Rubric

The authors define a rubric that distinguishes between Clickability (legitimate, bounded engagement) and Clickbait (disproportionate, misleading amplification).

• Key Attributes: Information gap control, emphasis intensity, emotional framing, salience allocation, structural emphasis, referential clarity, reader addressing, narrative structure, conceptual framing, and interrogative form.
• Differentiation: For each attribute, the rubric defines a "Clickability Interpretation" (e.g., Partial disclosure that invites interest) vs. a "Clickbait Interpretation" (e.g., Deliberate withholding of essential facts).

B. Synthetic Dataset Construction

Since real-world datasets often lack fine-grained attribute labels, the authors generated a synthetic corpus:

1. Source: Neutral headlines from the ISOT Fake News Dataset (True.csv subset from Reuters).
2. Generation: Using GPT-4o-mini, they rewrote neutral headlines by activating specific "clickbait-oriented" tactics (e.g., Curiosity Gap, Exaggeration) while preserving factual content.
3. Structure: Each sample includes the headline, a binary clickbait label, and a sparse vector indicating which specific engagement attributes were activated.

C. Auxiliary Guide Models

Two BERT-based models are trained on the synthetic dataset to provide control signals during decoding:

1. Clickbait Scoring Model (Negative Guidance): A binary classifier that predicts the probability of a text prefix being clickbait. It penalizes excessive stylistic amplification.
2. Engagement Attribute Model (Positive Guidance): A multi-label classifier that predicts the presence and strength of specific engagement attributes (e.g., "Curiosity Gap"). It encourages the realization of target rhetorical tactics.

• Training Nuance: Both models are trained on token-level prefixes (not just full sentences) to provide real-time feedback during the LLM's generation process. A length-aware loss function is used to down-weight very short, uninformative prefixes.

D. FUDGE-Based Controlled Decoding

The system uses the FUDGE framework to steer a base LLM (e.g., GPT-4o-mini) at inference time without modifying its weights.

• Mechanism: At each decoding step, the base LLM proposes candidate tokens. The auxiliary models evaluate the extended prefix (current text + candidate token).
• Scoring Function: The final probability for a token $y$ is calculated as:
  $$\log p(y|x_{1:t}) \propto \log p_{LLM}(y|x_{1:t}) + \lambda_{pos} S_{pos}(x_{1:t} \oplus y) - \lambda_{neg} S_{cb}(x_{1:t} \oplus y)$$
  • $S_{pos}$: Weighted sum of desired engagement attributes.
  • $S_{cb}$: Clickbait score (penalty).
  • $\lambda_{pos}, \lambda_{neg}$: Tunable weights to balance engagement vs. restraint.

3. Key Contributions

1. Reframing Clickbait: Proposes that clickbait is not a distinct category but a graded consequence of disproportionate engagement attribute activation.
2. Controllable Rubric: Defines an interpretable space of linguistic dimensions separating legitimate optimization from manipulative exaggeration.
3. Dual-Guidance Framework: Introduces a novel inference-time control mechanism using positive guidance (to boost specific engagement tactics) and negative guidance (to suppress clickbait), avoiding the need for expensive RL fine-tuning.
4. Prefix-Aware Training: Develops auxiliary models capable of evaluating partial generations, enabling real-time steering of LLMs.
5. Empirical Validation: Demonstrates that the system can generate headlines that are more engaging than neutral ones but significantly less sensationalist than unconstrained rewrites.

4. Results

• Scoring Models:
  • The Clickbait Scoring Model achieved an AUROC > 0.99 on the synthetic test set, indicating near-perfect discrimination between neutral and clickbait variants.
  • The Engagement Attribute Model successfully identified specific tactics. While some attributes (e.g., Curiosity Gap, Lists/Superlatives) were easily detected, others (e.g., Sensationalism) showed higher confusion, reflecting the natural overlap of rhetorical strategies.
• Qualitative Rewriting:
  • Unconstrained Rewriting ($\lambda_{neg}=0$): Produced headlines with high engagement but frequent exaggeration, implied secrecy, and emotional distortion (e.g., "What is behind...?" becoming "The shocking truth behind...").
  • Controlled Rewriting ($\lambda_{neg} > 0$): Produced headlines that retained the intended rhetorical tactic (e.g., curiosity) but remained semantically faithful and proportionate.
  • Example:
    • Neutral: "Venezuelan crisis spawns boom in gambling."
    • Unconstrained: "Venezuela's Desperate Situation Ignites a Frenzy of High-Stakes Bets." (Exaggerated)
    • Controlled: "Venezuela's Desperate Struggle Highlights Troubled Citizens' Frenzied Quest for Luck." (Engaging but restrained).

5. Significance and Future Work

• Significance: This work provides a principled, responsible AI approach to headline optimization. It proves that engagement and integrity are not mutually exclusive; rather, clickbait is a failure of constraint, not an inevitable byproduct of optimization. The framework is applicable to editorial tools, recommender systems, and content moderation.
• Future Directions:
  • Extending control to tactic-specific penalties (e.g., suppressing urgency inflation specifically).
  • Validating the approach with human judgments and real-world engagement metrics (CTR) to calibrate weights.
  • Applying the framework to longer-form content (summaries, social media posts).
  • Integrating the mechanism into multi-step generation loops for complex content composition.

In summary, the paper presents a robust, inference-time control framework that leverages the FUDGE paradigm to navigate the trade-off between attractiveness and truthfulness, offering a practical solution for ethical news headline generation.