Can user history override an LLM's politeness bias in reviews?

LLMs trained on web text and aligned via RLHF are systematically polite. They generate "I'm not a fan of this" rather than "this is terrible". For e-commerce review generation, this is a structural problem: users are dissatisfied with many items, and their reviews should be honestly negative. A polite LLM produces inappropriately positive reviews for items the user hated.

Review-LLM attacks this with two interventions. First, the prompt input aggregates the user's behavior sequence — item titles, prior reviews, and ratings — so the LLM has access to the user's review style and habits. This addresses the corpus-level pretraining problem (LLMs don't capture individual review style) by giving the model in-context examples of how the user writes. Second, the user's rating of the target item is included in the prompt as a satisfaction indicator. A 2-star rating tells the model the review should be negative, overriding the politeness default. Finally, the model is supervised-fine-tuned with the prompt-and-target pairs to lock in the personalization.

The result is a model that generates personalized negative reviews when the user is dissatisfied — outperforming closed-source LLMs that produce polite reviews regardless of the user's actual experience. The general principle: if the LLM's default behavior is misaligned with the task, providing both the personalized context (user history) and the explicit signal (rating as satisfaction proxy) lets fine-tuning override the default. Without both, fine-tuning alone struggles against the politeness bias.