If you have ever asked AI to write an email or a message, you have probably paid attention to a familiar defect: the results often look very common. Even with detailed instructions, large language models such as chatters or Gemini often struggle to reflect a person’s real tone or writing style. Apple believes that it has found a solution. In a new research paper to be presented at the International Conference on Machine Learning (ICML 2025), the researchers of Apple introduced a technique called prose, which is small for preference logic by looking and synthesizing examples. It is designed to help the AI ​​system learn directly from the user’s writing history, so future output can match the user’s natural tones and preferences.

What is prose and how it works?

Unlike older methods such as early engineering or reinforcement learning from human response, prose makes a unique and developed profile to write a person. It works in two major stages:

1. Itarative refinement: AI real user-compared to its generated draft with examples and adjusts its internal style details until the output does not resemble the user’s tone.

2. Consistency verification: The system then checks whether the identified writing preferences (such as “using small sentences” or “friendly and casual”) appear continuously in many samples. This ensures that AI does not depend on just one piece of writing.

The result is a self-teaching system that can tailor your future writing based on your overall writing behavior, not just a document or prompt.

While Apple did not name any of its products in the research paper, potential applications are clear. Apple Intelligence with the company’s increasing push-its next-gyne AI AI can allow apps such as all-facilities-melodies or notes to generate texts that make more sound like you.

Apple’s new Foundation Model Framework, recently announced, will allow developers to reach Apple’s local language model. This means that prose-style personalization can eventually power the writing tool inside a wide range of third-party apps.

Along with prose – Sath, Apple also introduced a new dataset called Plum – user shorter to learn preference from emails and memo. It replaces the previous dataset, prelood, and its purpose is to do better testing how well the AI ​​system can understand and repeat writing preferences.

When tested using the plum, prose defeated the existing privatization techniques. This improved a similar system called cipher than 33 percent and even showed better results than standard in-order learning (ICL) approaches, especially when used with advanced models such as GPT-4O. Interestingly, the combination of prose with ICL gave the best overall results – with a 9 percent improvement on ICL alone.

Why individually like prose matter

Apple’s prose research is not happening in separation, it follows a more important discovery from the company’s AI team, which raises serious questions about today’s big language model (LLM), especially when complex tasks are faced.

Title in a separate paper Confusion of thinkingWhich was released earlier this month, Apple’s researchers argue that even top-level models like chatgipt O3, Cloud 3.7 sonnet, and Dipsek-R1 may not be as intelligent as they seem. These models are often described as “large logic models” (LRMS) as they try to break the complex functions using step-by-step “chain”. But Apple’s study shows that this capacity can be more fragile than expected. To test it, the researchers created a custom environment – including the Tower of Hanoi, River Crossing, and Block Worlds – designed to increase the complexity of problems gradually. The objective was not only to test whether AI could get the correct answer, but can it logically make its way there.

The results were amazing. When the tasks were simple, traditional models without complex arguments strategies often performed better. At a moderate level of complexity, the logic-based model led briefly. But as – such as the complexity increased further, all models – the most advanced – has been experienced that researchers say “performance collapse”. In short, the AI ​​system which seemed smart or completely defeated, when the problems became very difficult – despite having enough computing power to continue.

This highlights why Apple’s prose approach is on time. If AI cannot “think” through really hard challenges, then very least, it should be able to do more human sound – learning how you write.