Apple recently launched the new App Store comment summary function with iOS 18.4. This function mainly uses the AI big model language, which can integrate multiple App Store comments into summary, allowing consumers to grasp the overall evaluation of App or game at a glance. Recently, Apple has rarely explained the operating details of the technology behind the App Store review summary in its own official Machine Learning Research blog.
How does the App Store review summary work?
Apple officially stated that comments and ratings are indispensable references for users to evaluate apps. The new App Store Review Summary feature mainly uses multi-stage artificial intelligence large language model (LLM), which can quickly compile massive comments into accurate comments of 100 to 300 characters, and can faithfully reflect users’ voices.
Four major challenges in AI generation summary: immediacy, diversity, accuracy and high quality
In the process of generating App Stroe summary, security, fairness, authenticity and practicality are core principles. However, if you want to integrate massive comments from the App Store, Apple actually faces many system design technical challenges.
- Immediateness: App comments will continue to change with new versions, feature updates and vulnerability corrections. The AI summary must be dynamically adjusted to reflect the latest user feedback.
- Diversity: User comments have different lengths, different styles, and large gaps in information volume. AI needs to capture key details while retaining diverse perspectives and levels.
- Accuracy: Not all comments focus on App content, some may be off topic or contain nonchalant information, and AI must have the ability to filter to ensure that the summary is credible.
- High Quality: Apple establishes rigorous evaluation standards, and professional teams repeatedly test the safety, aptitude, language style and practicality of the abstract.
App Store summary comment generation process is understood at one time
At the same time, Apple also shared this AI summary system based on LLM and fine-tuned by LoRA adapter, which successfully solved challenges such as dynamic and changeable comments and diverse content. The evaluation results show that the final output summary can truly reflect the user’s voice, while being safe, practical and high-quality language performance.
Below are five major processes and mechanisms in the generation process of the entire App Store:
- Preliminary filtering: AI will first automatically exclude comments containing spam, swear words or scam content.
- Insight extraction: LLM finely tuned by LoRA adapter refines each comment into a clear, single-themed insight sentence for easy cross-comment comparison.
- Dynamic Theme Modeling: LLM further categorizes insights, without a fixed taxonomy, and removes subject names with semantic duplications. The system will also judge the correlation between the topic and “App experience” or “App experience”, and give priority to organizing key topics such as functions, performance, and design.
- Topics and Insights: Automatically select representative insight sentences to summarize based on the topic’s popularity, balance, relevance and immediacy.
- Summary generation: Finally, another LLM finely tuned by LoRA adapter generates a natural, smooth, Apple-style summary based on selected insights, with lengths controlled between 100-300 characters. This model also combines professional abstract examples with human work as a reference, and further uses Direct Preference Optimization (DPO) mechanism to fine-tune it to improve language style and content performance.
Apple says this technology not only allows App Store users to browse and make decisions more efficiently, but also demonstrates the broad potential of LLM technology in large-scale user-generated content applications.
The official blog also announced that the team behind this technology includes experts from different fields, such as Sean Chao, Srivas Chennu, Yukai Liu, Jordan Livingston, etc., to work together to create the industry’s most advanced user review summary experience.
Ensure summary quality with manual review of automatic AI reviews
Of course, in the comments generated by AI, Apple also emphasized that thousands of App Store summary were finally reviewed by manual multiple reviews, which were evaluated from four aspects: security (whether the content is harmful), aptitude (whether it faithfully reflects the comments), language composition (grammar and brand style) and practicality (helping for download decisions). Each scoring standard has a strict review mechanism, and is supplemented by automated testing, allowing engineers to focus their human resources on the most needed improvements.
If you are interested in the application of AI technology to the high-traffic App Store platform to generate comments, you can go to Apple’s official blogCheck out the complete technical explanation. Finally, I would like to remind you that the current App Store generates AI comment function, which will only be displayed in the Apple Intelligence-supported national stores.
Further reading:
