As extra customers begin counting on AI for writing duties like e-mail drafts and doc summaries, one frequent frustration stays: the output usually sounds manner too generic. Even when fashions like ChatGPT or Gemini are given detailed prompts, they hardly ever nail a consumer’s particular person tone or voice with out a lot of guide tweaking. Apple is now proposing an answer.
In a brand new analysis paper (Aligning LLMs by Predicting Preferences from Consumer Writing Samples) to be offered on the Worldwide Convention on Machine Studying (ICML 2025) subsequent month, Apple researchers unveil PROSE, a way designed to assist giant language fashions higher infer and undertake a consumer’s distinctive writing preferences by studying instantly from their previous writing samples.
How PROSE works
The central concept behind PROSE (Preference Reasoning by Observing and Synthesizing Examples) is to maneuver past at the moment’s typical alignment strategies, like immediate engineering or reinforcement studying from human suggestions. As a substitute, the AI builds an inner and interpretable profile of the consumer’s precise writing fashion.
Moderately than requiring the consumer to manually present fashion guides or edit numerous AI drafts, PROSE works in two phases:
- Iterative Refinement: The AI repeatedly compares its personal generated responses with actual examples from the consumer, adjusting its inner “choice description” till it outputs one thing that intently matches the consumer’s writing.
- Consistency Verification: To keep away from fixating on only one instance, which could not be consultant of the consumer’s general writing fashion, the AI double-checks that any inferred choice (e.g., “use brief sentences” or “begin with a joke”) holds true throughout a number of writing samples.
Principally, PROSE builds a self-evolving fashion profile, checks it in opposition to a number of consumer examples, and makes use of that because the baseline for future generations.
Why this issues for Apple Intelligence
Whereas the paper doesn’t point out Apple services or products by title, the connection is apparent. As Apple pushes deeper into extra customized assistant options, strategies like PROSE might play a giant function in making Apple Intelligence write texts that really feel extra like every particular person consumer.
And with Apple now permitting builders to faucet instantly into its native fashions by way of the newly introduced Basis Fashions framework, it’s not onerous to think about a future the place any app might leverage a system-wide, deeply customized writing assistant to energy its personal writing instruments.
There’s a brand new benchmark, too
Within the research, Apple additionally introduces a brand new benchmark dataset referred to as PLUME (Preference Lincomes from User Emails and Memos) for evaluating writing-style alignment strategies like PROSE.
This replaces a earlier dataset (PRELUDE) and goals to repair frequent points with LLM personalization testing, like shallow choice definitions or non-representative duties.
Utilizing PLUME, the researchers in contrast PROSE to earlier approaches, corresponding to one other preference-learning methodology referred to as CIPHER (I do know. So many names and acronyms) and customary in-context studying (ICL) strategies.
The end result? PROSE outperformed CIPHER by 33% on key metrics and even beat ICL when paired with high-end fashions like GPT-4o.
Curiously, the paper additionally means that combining PROSE with ICL delivers one of the best of each worlds, with as much as a 9% enchancment over ICL alone.
The larger development: AI that adapts to you, and retains you coming again
The PROSE mission matches right into a broader AI analysis development: making assistants not simply smarter, however extra private. Whether or not that’s by way of on-device fine-tuning, choice modeling, or context-aware prompts, the race is on to shut the hole between generic LLM output and the distinctive voice of every consumer.
After all, true personalization additionally comes with enormous enterprise incentives, because it additionally units the stage for the last word platform lock-in. However that’s a topic for one more day.
FTC: We use earnings incomes auto affiliate hyperlinks. Extra.
