As increasing numbers of users depend on AI for tasks such as drafting emails and summarizing documents, one prevalent issue persists: the results often come off as overly generic. Even when advanced models like ChatGPT or Gemini receive comprehensive prompts, they seldom capture a user’s specific tone or voice without substantial manual adjustments. Apple is now introducing a potential fix.
In a new research publication set to be presented at the International Conference on Machine Learning (ICML 2025) next month, Apple researchers introduce PROSE, a method aimed at assisting large language models in more effectively discerning and adopting a user’s unique writing preferences by directly learning from their previous writing samples.
### How PROSE operates
The fundamental concept of PROSE (Preference Reasoning by Observing and Synthesizing Examples) is to advance beyond the standard alignment techniques of today, such as prompt engineering or reinforcement learning based on human feedback. Instead, the AI formulates an internal and interpretable profile of the user’s genuine writing style.
Rather than necessitating the user to manually supply style guides or modify numerous AI drafts, PROSE functions in two phases:
1. **Iterative Refinement**: The AI systematically compares its generated responses with actual examples from the user, fine-tuning its internal “preference description” until it yields an output that closely resembles the user’s writing.
2. **Consistency Verification**: To prevent fixation on a single example, which may not accurately reflect the user’s overall writing style, the AI verifies that any inferred preference (e.g., “use concise sentences” or “begin with humor”) remains consistent across various writing samples.
Essentially, PROSE creates a self-evolving style profile, tests it against multiple user examples, and uses that as a foundation for future outputs.
### Why this is significant for Apple Intelligence
Though the paper does not specifically name Apple products or services, the link is clear. As Apple delves further into personalized assistant functionalities, methods like PROSE could significantly impact Apple Intelligence’s ability to create texts that resonate more closely with each individual user.
With Apple also enabling developers to access its local models through the recently announced Foundation Models framework, envisioning a future where any application could utilize a system-wide, highly personalized writing assistant to enhance its writing capabilities is feasible.
### A new benchmark has emerged
In the research, Apple additionally presents a new benchmark dataset called PLUME (Preference Learning from User Emails and Memos) for assessing writing-style alignment techniques such as PROSE.
This supersedes an earlier dataset (PRELUDE) and aims to address common problems with LLM personalization assessments, such as superficial preference definitions or unrepresentative tasks.
Utilizing PLUME, the researchers juxtaposed PROSE with prior methodologies, including another preference-learning approach known as CIPHER and conventional in-context learning (ICL) techniques.
The outcome? PROSE surpassed CIPHER by 33% on crucial metrics and even outperformed ICL when paired with advanced models like GPT-4o.
Notably, the paper also indicates that merging PROSE with ICL offers the advantages of both approaches, resulting in up to a 9% enhancement over ICL alone.
### The overarching trend: AI that adapts to you and encourages ongoing engagement
The PROSE initiative aligns with a larger trend in AI research: enhancing assistants to be not only smarter but also more personalized. Whether it’s through on-device fine-tuning, preference modeling, or context-aware prompts, the race is on to bridge the gap between generic LLM outputs and the distinct voice of each user.
Naturally, genuine personalization carries substantial business incentives, as it lays the groundwork for ultimate platform lock-in. However, that is a topic for another discussion.
Read More
