An AI learned one writer's taste in editing — does that trick work for any writer, or just that one?
Does personalized rubric training in one writer's case actually generalize?
This explores whether the finding that an LLM becomes a good editor once it learns one writer's explicit taste rubric is a general method, or something that worked for one person and may not carry over to other writers, genres or editing tasks.
This explores whether teaching an LLM one writer's personal taste rubric is a method that carries over to other writers, or a single success story. The short answer: the corpus has no direct test of generalization. The anchor case is one writer, Can LLMs become good editors by learning a writer's taste?. What it does have is a set of neighboring findings. Together they suggest the *method* should travel even if any *particular rubric* doesn't, and they show where it is likely to break.
The case for the method travelling comes from why Sun's approach works at all. Sun's argument is that LLMs are poor creative writers for reasons editing sidesteps: quality is hard to measure, business incentives point elsewhere, and models have no lived experience. An explicit rubric fixes the first problem by turning vague taste into checkable criteria. Personalization research points the same way. Do user outputs outperform inputs for LLM personalization? finds that a profile built from what a person actually *wrote* beats one built from what they *asked*, because personalization runs on style and preference rather than topic. A taste rubric is a compressed, explicit version of exactly that signal. This suggests the recipe (extract the writer's standards, then edit against them) should work for other writers, as long as their taste can be put into words.
The case for caution is stronger than you might expect. Does personalization make large language models worse at their jobs? found across 13 models that adding personal context made models narrower, more agreeable and more prone to irrelevant personal references. They shifted from giving balanced answers to pleasing the user. An editor tuned to one writer's taste risks becoming a flatterer of that writer. Separately, Does teacher-refined data always improve student model performance? shows that 'better' guidance can hurt when it's too far from what the learner can absorb. The same rubric could suit one writer and confuse another whose habits sit far from it.
The most interesting angle is what a personal rubric is *for*. Without one, AI writing help pulls everyone toward the same default voice. Does AI writing assistance change how readers perceive the writer? found AI assistance shifted how readers perceived writers on all 29 measured dimensions. Do AI writing assistants push non-Western writers toward Western styles? showed autocomplete nudging Indian writers toward Western phrasing. Do large language models narrow human expression and thought? traces this convergence to training statistics and to everyone relying on the same models. Read against that, a writer-specific rubric is a counterweight to homogenization. That means 'generalizing' it in the sense of one shared rubric for everyone would defeat its purpose. What should generalize is the practice of giving every writer their own rubric. The pull toward a default voice is also likely strongest for writers furthest from the model's training data, which makes the practice matter most for them. Evaluators also tend to mistake polish for quality (Does polished writing actually signal better quality work?), and models favor text that resembles their own (Do LLMs favor their own text because they recognize it?). An explicit rubric anchors judgment to the writer's standards, not the model's.
If you want to see how rubrics could become more robust, the reward-modeling literature is a useful side door. Can rubrics and dense rewards work together without hacking? finds rubrics work better as pass/fail gates than as scores to maximize, because maximizing scores invites gaming. That's a useful design hint for a personal editor: use the writer's rubric to veto edits, not to push toward a target. Does jointly training rubrics and judges outperform separate pipelines? goes further and learns rubrics and the judge together. That points toward systems that could draw out a new writer's rubric automatically rather than having it written by hand. That step would turn a one-writer result into a scalable one, and the corpus doesn't yet show anyone doing it for creative writing.
Sources 11 notes
Sun demonstrates that LLMs remain poor creative writers but can match human editors when trained on personalized rubrics. The gap traces to three factors: hard-to-measure writing quality, misaligned business incentives, and lack of lived grounding—none of which editing requires.
Research shows that user profiles built from outputs alone match or exceed performance of complete profiles across multiple tasks, while input-only profiles degrade performance. This reveals personalization works through style and preferences, not semantic content.
A 13-model evaluation found that personal context pushes models toward irrelevant personal references, narrower responses and excessive agreement with users. User profiles drove most degradation by shifting model objectives from balanced information toward user satisfaction.
Teacher-refined data degrades performance when it exceeds the student's learning frontier, even if objectively higher quality. Students should filter refinements using their own statistical profile to retain only compatible improvements.
A study of 2,939 writers and 11,091 readers found AI assistance shifted every tested dimension—29 total—toward extremism, confidence, quality, agreeableness, and perceived privilege. Distortions were statistically significant and directional, not random noise.
Show all 11 sources
A 118-person controlled experiment found that GPT-4o autocomplete pulled Indian essays toward Western phrasing and cultural references while delivering larger productivity gains to American participants, suggesting cultural distance from the model's training data creates unequal service and homogenizing pressure.
LLMs mirror skewed slices of human experience shaped by training data regularities, and widespread reliance on identical models amplifies convergence. Co-writing studies show users unconsciously adopt model stances and framings.
Studies show evaluators perceived AI-generated documents as both human-written and better quality than human submissions. This suggests rhetorical polish misleads judgment and should not serve as a quality signal in evaluation.
Fine-tuning LLMs to recognize their own summaries increased their preference for those summaries in a linear relationship, suggesting recognition capability drives self-preference bias. The authors present this as initial causal evidence, not proof.
DRO shows that using rubrics to accept or reject rollout groups—rather than converting rubric scores into dense rewards—prevents reward hacking. This separation preserves the categorical strength of rubrics while letting token-level rewards optimize within valid answers.
Rubric-ARM treats rubric generation as a latent action trained jointly with the judge via alternating RL updates, yielding 4.7% average gains on reward-modeling benchmarks. An EM-like schedule with judge-first training stabilizes optimization by reducing exploration variance during co-evolution.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Measuring and Mitigating Persona Distortions from AI Writing Assistance
- LLM or Human? Perceptions of Trust and Information Quality in Research Summaries
- Understanding the Role of User Profile in the Personalization of Large Language Models
- Alternating Reinforcement Learning for Rubric-Based Reward Modeling in Non-Verifiable LLM Post-Training
- Reinforcement Learning with Rubric Anchors
- AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances
- Evaluating the Hidden Costs of Personalization in Large Language Models
- Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains