In the modern world, artificial intelligence has become a familiar companion for answering questions, from planning a trip to drafting an email. In the realm of health and nutrition, people increasingly turn to chatbots for advice on what to eat or how to lose weight. However, a significant gap remains between a generic answer and one that truly helps an individual. A recommendation that ignores a person's specific age, weight, medical history, and daily habits is fundamentally different from a personalized plan. While many current tools attempt to tailor advice by simply pasting a user's profile into a request, they often fail to truly understand the details before speaking. This leaves a critical structural gap: the computer sees the data, but it does not pause to analyze what that data means for the specific person asking the question.

Researchers at ITMO University in Saint Petersburg have developed a new method to bridge this gap, aiming to make nutritional advice safer and more useful. They call their approach PA-CoT, which stands for Profile-Adaptive Chain-of-Thought. Instead of rushing to generate an answer, this method forces the artificial intelligence to take a deliberate step back and analyze the user's profile as a separate, dedicated task. Think of it as a doctor who reads a patient's chart and writes down the key issues before speaking to the patient, rather than trying to figure out the diagnosis while talking. The researchers tested this idea against eleven other common ways of prompting AI, using a new set of 200 realistic nutritional consulting scenarios. Their findings show that by separating the analysis of the user's data from the generation of the advice, the system produces recommendations that are significantly more personalized and safer than those produced by standard methods.

The core of this new method lies in how it structures the thinking process. In traditional approaches, an AI receives a question and a block of text containing the user's profile, then attempts to answer everything at once. The researchers found that this often leads to the AI missing crucial details or offering generic advice that could be harmful. PA-CoT changes the workflow by breaking it into distinct stages. First, the system analyzes the user's profile to identify the most critical points that the answer must address. It looks at specific values, such as age, weight, and dietary goals, and determines exactly what the response needs to cover. This step creates a clear roadmap for the next stage. Only after this analysis is complete does the AI generate the actual response, using the roadmap to ensure every piece of relevant data is woven into the advice. Finally, a third stage acts as a safety check, reviewing the generated answer to ensure it contains no harmful recommendations and includes necessary warnings for vulnerable groups.

To test whether this extra step actually made a difference, the team created a new benchmark called QPA, which stands for Question-Profile-Answer. They built this dataset from thousands of real medical questions found in online forums, filtering them to keep only those related to nutrition. For each question, they extracted the user's profile details, such as their age, weight, and health goals, creating a structured set of 200 test cases. These cases were designed to mimic real-world situations where data might be incomplete, just as it often is in actual health apps. The researchers then ran their PA-CoT method against a wide range of other techniques, including standard prompting, methods that try to refine answers through repetition, and automated systems that optimize how questions are asked. Each response was scored by an advanced AI judge on four criteria: whether the advice was factually correct, whether it covered all parts of the question, how well it used the user's specific profile, and whether it was safe.

The results were clear and decisive. PA-CoT achieved the highest average score among all twelve methods tested. More importantly, it led the pack in the two areas that matter most for health advice: personalization and safety. The method scored significantly higher on personalization than its closest competitor, with a gap so large that the statistical confidence intervals did not overlap. This means the improvement was not a fluke but a consistent result. Similarly, it scored highest on safety, ensuring that the advice was free from harmful recommendations. The researchers found that other complex methods, which tried to optimize the prompt automatically or refine the answer through multiple iterations, did not perform as well. In fact, some of these more complicated approaches scored lower on personalization than simple methods. This suggests that adding complexity does not automatically lead to better results; instead, the key was the specific architectural choice to analyze the user's profile before generating the response.

One of the most striking findings was how the method handled the trade-off between being helpful and being safe. In many previous attempts, making an AI more personalized often made it riskier, as the system might try too hard to fit the user's data and accidentally suggest something dangerous. PA-CoT solved this by separating the tasks. The stage that creates the personalized advice focused entirely on using the user's data, while a separate stage focused entirely on safety. This allowed the system to be highly tailored without compromising on safety. For example, in a test case involving a teenager trying to lose weight, the system correctly identified that the user's age and weight were critical factors and generated a plan that respected those limits, while also flagging the need for professional medical supervision. In contrast, other methods often gave generic advice that ignored the user's age or failed to include necessary safety warnings.

The researchers also explored how much data was needed for the system to work effectively. They found that even when the user's profile was incomplete, with many fields marked as unknown, the method still performed well. It could identify which pieces of information were missing and, in a real-world product setting, could be designed to ask the user follow-up questions to fill those gaps. In the current study, the system worked with the fixed data available, but the architecture was built to handle the dynamic nature of real conversations. The cost of using this method was also considered; while it required more computer processing steps than a simple one-step answer, the extra cost was minimal compared to the significant gain in quality and safety.

This work highlights a fundamental shift in how artificial intelligence can be applied to sensitive fields like health. It demonstrates that the way a model is prompted to think is just as important as the model itself. By treating the user's profile not as a static block of text but as a subject for dedicated analysis, the researchers created a system that respects the individuality of the person asking for help. The study does not claim that this system replaces professional doctors or dietitians; the authors explicitly state that it is a research prototype and that real-world deployment would require validation by human experts. However, the results provide strong evidence that a structured, multi-step approach to reasoning can significantly improve the quality of AI-generated advice. As artificial intelligence becomes more integrated into daily life, methods like this offer a path toward tools that are not only smart but also thoughtful, safe, and truly helpful for the people who need them most.