Imagine working in a digital newsroom in Bangladesh. Your role is to read long news reports and write short, impactful headlines. But there’s a challenge. Certain words are off-limits. These restricted words vary with each article and are only provided at the time of writing. They may include overly common phrases, repetitive terms, or editorially sensitive words.

Now, imagine automating this task with AI.

In this challenge, an AI model has already been trained to generate headlines from news articles. However, during inference, it must generate a headline without using a specific list of restricted words, which are revealed only at inference time. This setup mirrors real-world editorial constraints where rules change dynamically.