Personalized Causal Graph Reasoning for LLMs: A Case Study on Dietary Recommendations
Zhongqi Yang, Amir M. Rahmani · 2025
Read Personalized Causal Graph Reasoning for LLMs: A Case Study on Dietary Recommendations by Zhongqi Yang, Amir M. Rahmani (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-00134.txt.
Introduction
Zhongqi Yang and Amir Rahmani propose a 2025 framework that helps large language models give advice shaped to one person. Their case is food advice for glucose control, where the same meal can affect people in unlike ways. Each user gets a causal graph built from data over time, showing how personal factors may drive a result.
When a query arrives, the model walks the graph, ranks paths by likely impact, and tests what might happen under other choices. The authors use counter checks to judge food ideas across three time windows after meals. They report lower glucose rise than prior methods, plus higher scores from model based judging of how personal the advice feels.
Catalog record
| CREATOR | Zhongqi Yang, Amir M. Rahmani |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-00134
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Personalized Causal Graph Reasoning for LLMs: A Case Study on Dietary Recommendations Zhongqi Yang, Amir M. Rahmani Abstract Large Language Models (LLMs) effectively leverage common sense knowledge for general reasoning, yet they struggle with personalized reasoning when tasked with interpreting multifactor personal data. This limitation restricts their applicability in domains that require context-aware decision-making tailored to individuals. This paper introduces Personalized Causal Graph Reasoning as an agentic framework that enhances LLM reasoning by incorporating personal causal graphs derived from data of individuals. These graphs provide a foundation that guides the LLM’s reasoning process. We evaluate it on a case study on nutrient-oriented dietary recommendations, which requires personal reasoning due to the implicit unique dietary effects. We propose a counterfactual evaluation to estimate the efficiency of LLM-recommended foods for glucose management. Results demonstrate that the proposed method efficiently provides personalized dietary recommendations to reduce average glucose iAUC across three time windows, which outperforms the previous approach. LLM-as-a-judge evaluation results indicate that our proposed method enhances personalization in the reasoning process. I Introduction Large Language Models (LLMs) have recently emerged as powerful tools for reasoning tasks across various domains [ 1 , 2 ] . Their ability to leverage extensive general knowledge has encouraged researchers to explore their applications within specialized fields, including healthcare [ 3 , 4 , 5 ] . In particular, nutrition-oriented food recommender services stand to benefit from LLM-based approaches, as these systems require accessible, interpretable, and accurate prediction of personalized physiological responses to nutrient intake [ 6 , 7 , 8 ] . Effective nutritional recommendations generated by LLM agents could assist individuals in managing chronic health conditions such as diabetes by acting as conversational agents based on personal metabolic responses to dietary choices [ 9 ] . Fig. 1: Comparison between a standalone LLM and the proposed Personalized Causal Graph Reasoning for dietary recommendations. The standalone LLM relies on generic reasoning and may provide risky advice, while our method utilizes a personal causal graph to assess individual metabolic responses for more precise recommendations. However, current LLMs and agents face challenges when reasoning over personal, context-specific data [ 10 , 11 , 12 ] . Existing methods rely primarily on broad, population-level knowledge derived from large-scale pretraining, which leads them to produce generic rather than personalized recommendations [ 13 , 14 , 15 ] . This limitation contrasts with human reasoning, where individuals inherently account for personal factors and contextual information when making health-related decisions [ 16 ] . For instance, in nutrient-based health interventions, the effectiveness of dietary changes depends on a combination of an individual’s metabolic history, underlying conditions, specific nutrient deficiencies, and general nutritional principles. As demonstrated in Figure 1 , standalone LLM often fails in such settings because they do not have a structured mechanism to reason over personal causal understanding of dietary effects from data [ 6 , 4 ] . To address this challenge, we introduce Personalized Causal Graph Reasoning that enhances LLMs’ personalized reasoning within an agentic framework. The framework constructs a personal causal graph for each user based on longitudinal health data to capture the unique user characteristics. LLMs then reason over this structured representation by dynamically exploring causal graphs and retrieving relevant external knowledge to generate personalized recommendations. Unlike conventional LLMs that process user queries under a generic reasoning paradigm, our approach provides a structured foundation that allows LLMs to perform personalized inference over explicit causal relationships. To evaluate the effectiveness of the proposed framework, we conduct a case study on dietary recommendations. Using a dataset comprising continuous glucose monitoring data, food intake logs, and physical activity records, we construct personal causal graphs that capture the relationship between nutrient intake and glucose regulation. The LLM utilizes these graphs to simulate dietary interventions and recommend foods that are expected to improve glucose stability. We propose counterfactual evaluation to assess whether the model’s recommended foods would have led to actual health improvements [ 17 ] . Additionally, we employ LLM-as-a-judge to assess whether our method improves the reasoning process by making it more personalized. This paper’s contribution is two-fold: • We introduce Personalized Causal Graph Reasoning that enables LLMs to perform personalized reasoning by incorporating causal graphs derived from personal data. • We evaluate the proposed framework through a case study on personalized dietary recommendations. To assess its effectiveness, we introduce a counterfactual evaluation method that estimates the potential glucose impact of LLM-generated food recommendations. II Related Works II-A LLMs and Reasoning Several techniques were proposed to elevate LLMs’ general reasoning tasks. Chain-of-Thought (CoT) as a classic prompting method enhances problem-solving by enabling the generation of intermediate reasoning steps [ 18 ] . Building upon CoT, approaches such as Tree of Thoughts (ToT) and Graph of Thoughts (GoT) have been proposed to further refine LLM reasoning in a more structured manner [ 19 , 20 ] . ToT allows models to explore multiple reasoning paths, while GoT models information as an arbitrary graph, to combine various reasoning paths into cohesive outcomes. By incorporating structured reasoning techniques, LLMs have demonstrated promising performance in decision-making tasks that require dietary knowledge [ 21 , 4 ] . Beyond prompting techniques, studies have explored iterative reasoning refinements. These include generating multiple reasoning paths and selecting the most consistent one, applying step-wise verification, and integrating feedback mechanisms to improve logical consistency [ 22 , 23 , 24 ] . Additionally, Gao et al. propose meta-reasoning, where LLMs dynamically select and apply different reasoning strategies based on the problem context [ 25 ] . The Reasoning on Graphs (RoG) synergizes LLMs with knowledge graphs to enable faithful and interpretable reasoning [ 26 ] . RoG employs a planning-retrieval-reasoning framework, where relation paths grounded in knowledge graphs are generated as faithful plans. These plans are then used to retrieve valid reasoning paths from the graphs. Their reliance on large-scale, population-level data limits the applicability in contexts requiring precise, personalized reasoning. This limitation arises because they primarily operate on unstructured text prompts and lack mechanisms to incorporate structured representations of personal information. Consequently, they struggle to model the intricate interplay of personal factors necessary for tailored decision-making. However, approaches like RoG offer promising directions to overcome these challenges by using graphs as a bridge to connect personal data to LLM reasoning. II-B Nutrition-Oriented Recommendations Nutrition recommendation systems aim to provide dietary advice tailored to individual health needs. Traditional systems often use collaborative filtering, leveraging user interactions and preferences to generate suggestions [ 27 , 28 , 29 ] . However, they fail to capture the complex causal relationships between dietary factors and health outcomes and struggle to adapt dynamically to changes in an individual’s health status [ 8 , 30 ] . We focus on recent advancements that have explored the performance of LLMs on personalized dietary recommendations [ 31 , 32 , 6 ] . For instance, ChatDiet combines personal and population models to generate tailored food suggestions [ 6 ] . It employs Retrieval-Augmented Generation (RAG) to retrieve triplets from a pre-constructed causal graph, then structures them into prompts that guide the LLM in recommendation generation. While this approach enhances personalization, it relies on a fixed pattern of retrieving specific triplets to inform the LLM’s responses.