Baidu — Why Your AI-Written Headline Sounds the Same to Everyone

Ask an LLM for the single best headline and you get a generic one that pleases the average user and nobody in particular. Baidu's GESE generates a diverse set, then picks per user in real time, lifting CTR 2.57%.

Key takeaways:

• “Best on average” means generic. Training an LLM to produce one best headline leads to mode collapse: safe patterns that miss users with more specific tastes.
• Generate to explore. Stage one trains the LLM with GSPO and a hierarchical reward to produce a set of headlines covering as many user interests as possible.
• Select to exploit. Stage two is a lightweight, real-time selector that picks the best candidate for each user from fresh contextual signals, keeping the heavy LLM out of the serving path.
• Results. On a feed with over 100 million daily active users, GESE lifted CTR 2.57% and dwell time 0.87%. The dwell gain suggests it isn't just clickbait.
• Our take. Explore-exploit isn't new; tying RL-driven generation to a live A/B test is the contribution.

Chapters:

00:00 Catering for 100 million guests

01:31 The business problem

03:03 Headlines frame the content

03:48 Why fine-tuning goes generic

06:05 Explore, then exploit

06:52 A reward for diversity

09:55 The real-time selector

11:28 Keeping LLMs off the hot path

12:14 Results at 100 million users

13:47 Why dwell time matters

15:18 How novel is it?

16:49 Comparisons with other work

19:08 Lessons for smaller teams

20:41 Recap

Sources:

• Generate to Explore, Select to Exploit: Aligning LLM-based Headline Generation with Personalized Recommendation (Chen et al., Baidu; arXiv, Sep 2026)

Related episode: E12: Netflix — One Model, One Pass, Your Whole Homepage

Enjoyed it? Follow Data Science in the Wild (one real production ML/AI system or paper per episode, hosted by Peter Liu) and rate it on Apple Podcasts or Spotify.

Also from Peter: The GenAI Evolution Atlas, Wednesday AI deep dives and Friday news roundups.

Made with AI-generated voices; curated by Peter Liu from the sources above.