Paper deep dive
ShanLiangRen: A Nutrition Agent for Personalized Daily Meal Planning
Miao Xie, Xiao Zhang, Yuan Wang, Ruixin Zhu, Chunli Lv
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Dietary nutrition planning plays an important role in chronic disease management and maintaining a healthy body. In applications, it must simultaneously satisfy personalized constraints and reasonable multidimensional nutritional goals. These two aspects often conflict, and user constraints evolve with feedback, resulting in a substantial gap between generic guidelines and executable plans. To bridge this gap, we first propose the personalized fully quantified multiobjective dietary planning problem (MDP). To tackle MDP, we develop a nutrition agent, ShanLiangRen. The system first transforms dietary specifications, nutrient data, user attributes and natural language requirements into an individualized constrained planning instance. It then employs an exact retrieval-augmented generation method to shrink the feasible candidate set from a large scale ingredient and recipe space. Finally, it adopts a refinement guided by Pareto principles, where an LLM iteratively revises candidate plans under deterministic nutrition computation and feedback from constraint verification. The system outputs fully quantified meal plans with explicit ingredients and portion sizes, together with reports on nutrition compliance that show constraint satisfaction and nutrient interval attainment. We have released the system online as a WeChat Program, ShanLiangRen. A demo video is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2610.07886v1
- Canonical: https://arxiv.org/abs/2610.07886v1
Trouble viewing inline? Open PDF directly ā
Full Text
33,088 characters extracted from source content.
Expand or collapse full text
ShanLiangRen: A Nutrition Agent for Personalized Daily Meal Planning Conference: Proceedings of the 35th ACM International Conference on Information and Knowledge Management; November 07ā11, 2026; Rome, ItalyProceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM ā26), November 07ā11, 2026, Rome, ItalyDOI: 10.1145/3799682.3840265ISBN: 979-8-4007-2539-5/2026/11CCS: Information systems Information retrievalCCS: Information systems Recommender systems Miao Xie email: 0520shui@163.com Affiliation: College of Information and Electrical Engineering, China Agricultural University, Beijing, China , Xiao Zhang email: B20243080802@cau.edu.cn Affiliation: College of Information and Electrical Engineering, China Agricultural University, Beijing, China , Yuan Wang email: wy929_cs@163.com Affiliation: College of Information and Electrical Engineering, China Agricultural University, Beijing, China , Ruixin Zhu email: zhuruixin07@126.com Affiliation: Key Laboratory of Precision Nutrition and Food Quality, China Agricultural University, Beijing, China and Chunli Lv email: lvcl@cau.edu.cn Affiliation: College of Information and Electrical Engineering, China Agricultural University, Beijing, China Ā© c Abstract. Dietary nutrition planning plays an important role in chronic disease management and maintaining a healthy body. In applications, it must simultaneously satisfy personalized constraints and reasonable multidimensional nutritional goals. These two aspects often conflict, and user constraints evolve with feedback, resulting in a substantial gap between generic guidelines and executable plans. To bridge this gap, we first propose the personalized fully quantified multiobjective dietary planning problem (MDP). To tackle MDP, we develop a nutrition agent, ShanLiangRen. The system first transforms dietary specifications, nutrient data, user attributes and natural language requirements into an individualized constrained planning instance. It then employs an exact retrieval-augmented generation method to shrink the feasible candidate set from a large scale ingredient and recipe space. Finally, it adopts a refinement guided by Pareto principles, where an LLM iteratively revises candidate plans under deterministic nutrition computation and feedback from constraint verification. The system outputs fully quantified meal plans with explicit ingredients and portion sizes, together with reports on nutrition compliance that show constraint satisfaction and nutrient interval attainment. We have released the system online as a WeChat Program, ShanLiangRen. A demo video is available at https://w.youtube.com/watch?v=652OtY5VlGA. Keywords: Dietary Planning, Retrieval-Augmented Generation, Agent ā c-license: by 1. Introduction Unhealthy diets have become an important driving factor of the global burden of chronic diseases: a systematic analysis covering 195 countries estimated that in 2017 about 11 million deaths were associated with dietary risks and caused about 255 million Disability-Adjusted Life Years (DALYs) (Afshin et al., 2019). Meanwhile, an increasing number of studies indicate that more scientific and individualized dietary interventions can affect metabolic health and disease risk(Corbin et al., 2025; Xie et al., 2026a). Existing dietary-planning studies follow two main directions. Health-oriented studies target chronic-disease risk factors, such as obesity and abnormal glucose levels, and develop intervention schemes or clinical nutrition management workflows centered on nutritional goals and food combinations (Hassapidou et al., 2023; Mehta et al., 2024; Karjoo et al., 2025). Computational studies model menus or recipes as combinatorial optimization or recommendation problems under nutritional and preference constraints, enabling the automatic generation of executable diet plans (Li and Zaki, 2020; Tian et al., 2022a; Tian et al., 2022b; Islam et al., 2023; Sahin and Aytekin-Sahin, 2024; Ataguba and Orji, 2025). However, two limitations remain: (1) personalized constraints evolve with usersā physiological states and often conflict with reasonable nutritional goals, making guideline-based plans difficult to follow and execute at the individual level; and (2) natural-language requirements lack a unified representation and must be reliably normalized into executable constraints and goals, while open-ended dialogue often drifts toward general nutrition Q&A with limited planning actionability (Kim et al., 2024; Ataguba and Orji, 2025). Recent studies have explored integrating LLMs with structured nutrition knowledge or multiobjective optimization to support personalized and explainable dietary recommendations (Gao et al., 2025; Zhang et al., 2025). Accordingly, we formulate the fully quantified multiobjective dietary planning problem (MDP). Given dietary specifications and user information, MDP aims to dynamically generate a plan suited to the userās physical condition, quantify its ingredients and portions, and address multidimensional targets beyond calories and macronutrients. Solving MDP involves three coupled challenges. First, the system must formulate an individualized planning instance by converting general dietary specifications into nutrient target intervals based on the userās physical attributes, while incorporating stable restrictions and daily preferences as requirements or priorities. Second, it must construct feasible meal candidates from structured food and recipe tables with deterministic nutrient composition data, ensuring reliable evidence for ingredients, portions, and nutrient values. Third, it must refine feasible candidates across multiple nutritional objectives and personalized preferences, using deterministic nutrient computation to verify compliance with the target intervals. To address MDP, we develop ShanLiangRen, an interactive system for personalized daily dietary planning. It focuses on constructing feasible candidates from a large ingredient space and producing executable solutions under multiple objectives through two key designs. First, an exact RAG mechanism based on subgraph matching reduces the feasible search space and generates candidate plans. It represents dietary specifications and user constraints as retrieval filters, abstracts meal structure, category coverage, and combination principles into a query subgraph, and retrieves combinations satisfying user requirements through exact structural matching over the knowledge graph. Second, a Pareto guided multiobjective refinement strategy balances constraint satisfaction and nutritional goals. Through a feedback loop of deterministic nutrient computation and constraint verification, it evaluates and optimizes the candidates, improving nutritional objectives while satisfying personalized constraints, and outputs the recommended plan. This demo makes three contributions: (1) we define the personalized fully quantified multiobjective dietary planning problem (MDP); (2) we design an exact retrieval and Pareto guided refinement pipeline with deterministic nutrient verification; and (3) we develop a system for personalized daily meal planning. 2. Problem Definition We represent a dietary planning task as an instance I=(S,T,u)I=(S,T,u). The planning instance is grounded in external nutritional knowledge rather than generated nutrition claims. Specifically, S=(G,P)S=(G,P) denotes the dietary specifications, where G is derived from the Chinese Dietary Guidelines (2022) and P follows the Chinese Food Guide Pagoda (2022) (Chinese Nutrition Society, 2022). T is the ingredient nutrient table built from the China Food Composition Tables (Yang Yuexin, 2018; Yang Yuexin, 2019), and u=(uattr,ureq)u=(u_attr,u_req) is the user information. Here, uattru_attr describes the userās basic physical signs and health attributes, and urequ_req describes the userās dynamic requirements. Therefore, the LLM is used to normalize user requirements and refine candidate plans, while nutrient values, target intervals, and constraint verification are grounded in external specifications and structured data. The system outputs a fully quantified daily plan x=xmā£māMx=\x_m mā M\, where M is the set of meals, and each xmx_m gives the dishes/ingredients and their portion sizes. A feasible solution must satisfy two types of hard constraints: xā§Cuā(ureq)x C_u(u_req) and xā§Cqualā(S)x C_qual(S), corresponding respectively to the personalized requirements that the user must satisfy and the qualitative healthy structural principles that the specifications must follow. Based on T, we compute the daily intake nutrition vector nā”(x)=Computeā”(x,T)n(x)=Compute(x;T), and derive the individualized nutrition recommendation set Rā”(uattr,S)=Ļā”(uattr,S)R(u_attr,S)=Ļ(u_attr,S) from uattru_attr and S, based on which we construct the nutrition objective set Jnutā(S,uattr)J_nut(S,u_attr). Here Cuā(ureq)C_u(u_req) mainly covers individual hard restrictions such as allergies, taboos, meal structure requirements, and explicit daily preferences, while Cqualā(S)C_qual(S) captures dietary structure principles derived from guidelines such as category coverage and meal balance. For each jāJnutā(S,uattr)jā J_nut(S,u_attr), Rā”(uattr,S)R(u_attr,S) provides its reasonable intake interval [Lj,Uj][L_j,U_j]. To characterize the deviation degree of plan x on objective j relative to this interval, we define the interval deviation: (1) djā(x)=maxā”0,Ljānjā(x),njā(x)āUjmaxā”Īµ,UjāLj,d_j(x)= \0,\,L_j-n_j(x),\,n_j(x)-U_j\ \ ,\,U_j-L_j\, where njā(x)n_j(x) denotes the intake value of plan x on nutrition objective j, and Īµ is a very small constant. Under the premise of satisfying the hard constraints, the goal is to make all nutrition objectives fall into their reasonable intervals as much as possible, and minimize the overall deviation: (2) minLā”(x) \ \ L(x) =ājāJnutā(S,uattr)wjā djā(x) = _jā J_nut(S,u_attr)w_jĀ· d_j(x) s.t. .t. xā§Cu(ureq),xā§Cqual(S). x C_u(u_req), x C_qual(S). where wjw_j denotes the importance weight of nutrition objectives. 3. System Design Overview. Given an MDP instance, ShanLiangRen follows a five module online pipeline (Fig.1): (1) individual constraint and objective formulation, (2) exact retrieval for feasible space reduction, (3) quantified candidate generation, (4) deterministic nutrition evaluation, and (5) Pareto guided refinement with explanation. Implementation. ShanLiangRen employs GPT-5.4 as its LLM backbone. The LLM is responsible for requirement normalization, candidate proposal, and targeted plan repair. Figure 1. Overall architecture of the ShanLiangRen system. (1) Dynamic Constraint Extraction and Task Structuring. The system first receives user information u=(uattr,ureq)u=(u_attr,u_req). The LLM performs structured extraction on urequ_req to form executable user constraints Cuā(ureq)C_u(u_req), covering hard requirements such as taboos/allergies, taste preferences, and meal/course structure (e.g., number of meals, number of dishes per meal, no repetition)(Yao et al., 2023; Schick et al., 2023). Meanwhile, the system constructs the individualized nutrition objective set Jnutā(S,uattr)J_nut(S,u_attr) based on uattru_attr and the specifications S, and generates a recommended intake interval [Lj,Uj][L_j,U_j]. This step uniformly maps user inputs to the formal objects in the problem definition, providing a consistent interface for retrieval reduction and plan generation. In our demo, ShanLiangRen adopts a controlled, plan based interaction design. Stable user attributes and stable restrictions are collected through structured fields, while daily preferences and revision requests can be expressed in natural language with respect to the current meal plan. This design preserves personalized expressiveness while preventing the interaction from drifting into open-ended nutrition consultation. (2) Exact RAG for Feasible Space Reduction. To avoid combinatorial search over the full library, the system adopts exact RAG based on subgraph matching to reduce the candidate space to RuR_u. (Lewis et al., 2020; Guu et al., 2020; Khattab and Zaharia, 2020; Xie et al., 2026b; Zhang et al., 2026) Specifically: (1) exact filtering of hard constraints: it first removes items that violate Cuā(ureq)C_u(u_req) in the ingredient/recipe library (e.g., forbidden ingredients, infeasible meal structures), and combines the qualitative specification constraints Cqualā(S)C_qual(S) for structural elimination (e.g., basic feasibility of category coverage/combination templates). For example, a requirement such as avoiding a specific ingredient, including a vegetable category in lunch, and preventing repeated staple foods can be converted into exclusion predicates at the ingredient level, category coverage constraints, and meal structure relations. Exact retrieval means that candidates violating these hard predicates or structural relations are excluded before LLM generation, so the LLM only operates on a user relevant feasible subspace. (2) subgraph matching retrieval: it abstracts āmeal structure/category coverage/combination principlesā as a query graph, and performs exact matching retrieval under structural constraints on the filtered set to obtain the candidate subspace most relevant to the current user context, yielding RuāDR_u D. After this reduction, subsequent LLM generation and refinement are performed only within RuR_u. Unlike conventional RAG that primarily retrieves semantically relevant textual contexts, our exact RAG retrieves structured meal candidates that explicitly satisfy query graph relations. If no feasible candidate satisfies all hard constraints, ShanLiangRen does not silently relax safety-critical requirements. Instead, the system reports the infeasible constraint combination and requests revision of lower-priority daily preferences before replanning. (3) Candidate Generation. Given RuR_u, the LLM generates a diverse set of candidate daily plans x(1),ā¦,x(K)\x^(1),ā¦,x^(K)\ (Wang et al., 2023). Each candidate explicitly specifies meal partitioning and ingredient/dish portion sizes, ensuring āfully quantified executabilityā. Since the generation is constrained by RuR_u, candidate plans inherit the retrieved feasibility conditions and are easier to verify and repair than unconstrained LLM outputs. The generation stage enforces structural hard requirements and promotes diversity for trade-off selection. Deterministic Evaluation. For each candidate x, the evaluator computes the daily intake vector nā”(x)=Computeā”(x,T)n(x)=Compute(x;T) based on the element table T, and checks whether hard constraints are satisfied (xā§Cuā(ureq)x C_u(u_req) and xā§Cqualā(S)x C_qual(S)). Then, according to the recommended interval [Lj,Uj][L_j,U_j] of each objective, the system computes the deviation djā(x)d_j(x) and aggregates it to obtain the overall deviation Lā”(x)L(x). The resulting deficit, excess, and interval attainment status are used both for user explanation and for the next refinement round. The LLM serves as a semantic proposal operator that combines retrieved ingredients according to natural-language preferences and produces targeted modifications while preserving unaffected constraints. (4) Pareto Guided Refinement with LLM. Unlike traditional Pareto optimization, we agentize the candidate generation and repair operators: the LLM performs iterative āminimal modificationā driven by deterministic evaluation feedback, rather than using handcrafted search operators (Meng et al., 2022). The system maintains an evaluated candidate pool and performs dominance filtering based on the deviation vector dā(x)=djā(x)d(x)=\d_j(x)\ (the smaller the better): it keeps candidate solutions that are no worse on all objective deviations and strictly better on at least one objective deviation, retains candidates reflecting different trade-offs, and removes dominated solutions that are not better on any considered dimension. Then, the system feeds the interval deviation feedback produced by the evaluator into the LLM, and the LLM performs āminimal-modificationā repairs, entering the next round until reaching an iteration limit or the improvement becomes stable. (5) Recommendation and Explanation. After the candidate pool converges, the system outputs: (1) the recommended plan xāx , selected from the non-dominated set according to the userās preference priorities; (2) an explainable report, showing the satisfaction of key hard constraints, nutrition computation results nā”(x)n(x), interval attainment for each objective (whether it falls into [Lj,Uj][L_j,U_j]) and the corresponding deviation djā(x)d_j(x)/deficit or excess. Figure 2. Interface of the ShanLiangRen system. Backend Knowledge Base. ShanLiangRen maintains a structured ingredientārecipe knowledge base in the backend, and uses the ingredient nutrition element table as the computation basis of the deterministic evaluator(Tian et al., 2022b). The knowledge base contains structured information including recipe entities, ingredient entities, ingredient category hierarchies, and macro and micro nutrient attributes, enabling the system to abstract requirements such as āmeal-type structure/category coverage/combination principlesā into query subgraphs and to perform exact matching retrieval within the candidate subspace. Data Scale: the current backend includes 1,550,1501,550,150 recipes and 64,12764,127 ingredients; each ingredient is annotated with 64 nutrient attribute dimensions, and a category hierarchy is provided. The backend knowledge base supports two different functions: structured retrieval and deterministic verification. Recipe and ingredient entities are used to construct candidate meal combinations under restriction and category coverage constraints, while the nutrient table is used to compute the actual intake values of generated plans. 4. System in Action As shown in Fig. 2, ShanLiangRen has been deployed as an online WeChat Mini Program, where users can access the system, submit dietary requirements, and receive plan based dynamic updates (Park et al., 2023; Chen et al., 2024): it stores stable restrictions as stable constraints, and incrementally re-plans when users revise daily needs via structured options and comments on the current plan. On the output side, the system generates a fully quantified daily plan with three meals, refined to dishes/ingredients and portion sizes, and performs deterministic nutrition computation based on the ingredient element table, providing interval comparisons for indicators such as energy and macronutrients as well as attainment verification of requirements from dietary specifications, and finally presents an executable personalized daily plan in the form of ārecommended plan + explainable verification reportā. In the demo, the user first fills in basic physical signs and stable preferences (Fig.2.2), and inputs daily requirements (Fig.2.1); the system immediately parses them into structured hard constraints and generates a fully quantified three-meal daily plan (Fig.2.3). Then, the interface shows nutrition computation results and specification attainment verification (Fig.2.4), and the user can choose among the recommended plan and several trade-off alternative plans. If the user temporarily adjusts requirements (e.g., taboos or meal structure changes), the system triggers fast incremental replanning and synchronously updates the explainable verification report. 5. Evaluation We evaluate ShanLiangRen on 300300 distinct planning instances covering common personalized requirements, including ingredient taboos, dietary preferences, meal-structure constraints, and nutrition goals. We compare ShanLiangRen with direct LLM generation and constrained prompting, where all methods receive the same user profiles and daily requirements. All generated plans are manually verified by a Ph.D. in nutrition based on the China Food Composition Tables (Yang Yuexin, 2018; Yang Yuexin, 2019), and the verification results are used to compute five metrics. PCS is the percentage of satisfied individual hard constraints. EPC is the percentage of plans with complete dish, ingredient, and portion information. NIA is the percentage of checked nutrient objectives whose verified values fall within the recommended intervals. Plans lacking sufficient portion information are marked as incomplete, and their corresponding nutrient objectives are treated as unverifiable. AND is the average normalized nutrient deviation Lā”(x)L(x) in Eq. (2), where lower is better. Latency is the average end-to-end generation time in seconds. Table 1. Overall comparison on 300 user planning cases. Method PCS NIA EPC AND Lat. (%)ā (%)ā (%)ā (norm.)ā (s)ā GPT-5.4 87.5 60.4 85.9 0.17 46 Gemini-3 75.0 55.2 80.2 0.25 19 Doubao-seed 2.0 87.5 61.1 79.9 0.18 27 ShanLiangRen 100.0 78.9 100.0 0.09 61 As shown in table 1, ShanLiangRen achieves 100.0%100.0\% PCS and EPC, with the highest NIA of 78.9%78.9\% and the lowest AND of 0.090.09, showing full individual constraint satisfaction, complete executable portion information, and lower overall nutrient deviation under full nutrition verification. GPT-5.45.4 and Gemini-33 provide partially complete portions but show lower PCS and NIA, while Doubao-seed 2.02.0 performs competitively on NIA but has weaker portion completeness.Although ShanLiangRen incurs a moderate latency overhead, the additional time is acceptable for daily meal planning. 6. Conclusion We presented ShanLiangRen, an interactive nutrition agent for personalized daily meal planning. We first defined the personalized fully quantified multiobjective dietary planning problem (MDP), then designed an exact retrieval and refinement pipeline guided by Pareto principles with deterministic nutrient verification, and finally developed an interactive system that outputs executable meal plans with portion sizes and compliance reports. The evaluation shows that ShanLiangRen improves constraint satisfaction, nutrient attainment, executability, and verifiability over direct LLM generation, at the cost of acceptable additional latency. Acknowledgements. This work was supported by the China Agricultural University āYoung Researcherā Start-up Fund No. QNYJY2024144 and the Visiting Scholar Program of the China Scholarship Council (CSC) No. 202506350123. References Afshin et al. (2019) A. Afshin, P. J. Sur, K. A. Fay, L. Cornaby, and G. Ferrara Health effects of dietary risks in 195 countries, 1990ā2017: a systematic analysis for the global burden of disease study 2017. The Lancet 393 (10184), p. 1958ā1972. External Links: ISSN 0140-6736, Document Cited by: Ā§1. Ataguba and Orji (2025) G. Ataguba and R. Orji Exploring large language models for personalized recipe generation and weight-loss management. ACM Trans. Comput. Healthcare 6 (2). Cited by: Ā§1. Chen et al. (2024) W. Chen, Y. Su, J. Zuo, C. Yang, C. Yuan, C. Chan, H. Yu, Y. Lu, Y. Hung, C. Qian, Y. Qin, X. Cong, R. Xie, Z. Liu, M. Sun, and J. Zhou AgentVerse: facilitating multi-agent collaboration and exploring emergent behaviors. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024, External Links: Link Cited by: Ā§4. Chinese Nutrition Society (2022) Chinese Nutrition Society Chinese resident dietary guidelines (2022). Peopleās Medical Publishing House, Beijing. External Links: ISBN 978-7-117-31404-6 Cited by: Ā§2. Corbin et al. (2025) K. D. Corbin, D. Igudesman, S. R. Smith, K. Zengler, and R. Krajmalnik-Brown Targeting the gut microbiotaās role in host energy absorption with precision nutrition interventions for the prevention and treatment of obesity. Nutrition Reviews 83 (10), p. 1928ā1943. Note: PMID: 40233201; PMCID: PMC12422011 External Links: Document Cited by: Ā§1. Gao et al. (2025) F. Gao, X. Zhao, D. Xia, Z. Zhou, R. Yang, J. Lu, H. Jiang, C. Park, and I. Li HealthGenie: a knowledge-driven llm framework for tailored dietary guidance. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management, p. 6639ā6643. External Links: Document, Link Cited by: Ā§1. Guu et al. (2020) K. Guu, K. Lee, Z. Tung, P. Pasupat, and M. Chang REALM: retrieval-augmented language model pre-training. CoRR abs/2002.08909. External Links: Link, 2002.08909 Cited by: Ā§3. Hassapidou et al. (2023) M. Hassapidou, A. Vlassopoulos, M. Kalliostra, E. Govers, H. Mulrooney, L. Ells, X. R. Salas, G. Muscogiuri, T. H. Darleska, L. Busetto, V. D. Yumuk, D. Dicker, J. Halford, E. Woodward, P. Douglas, J. Brown, and T. Brown European association for the study of obesity position statement on medical nutrition therapy for the management of overweight and obesity in adults developed in collaboration with the european federation of the associations of dietitians. Obesity facts 16 (1), p. 11ā28. External Links: Document, ISSN 1662-4025 Cited by: Ā§1. Islam et al. (2023) T. Islam, A. R. Joyita, Md. G. R. Alam, M. Mehedi Hassan, Md. R. Hassan, and R. Gravina Human-behavior-based personalized meal recommendation and menu planning social system. IEEE Transactions on Computational Social Systems 10 (4), p. 2099ā2110. External Links: Document Cited by: Ā§1. Karjoo et al. (2025) S. Karjoo, A. Braglia-Tarpey, A. P. Chan, A. G. Ayala GermĆ”n, R. E. Herdes, N. Pai, D. Sierra-Velez, B. Whitehead, R. E. Quiros-Tejeira, and D. Duro Evidence-based review of the nutritional treatment of obesity and metabolic dysfunction-associated steatotic liver disease in children and adolescents. Journal of Pediatric Gastroenterology and Nutrition 81 (3), p. 485ā496. Note: Epub 2025 Jun 17. PMID: 40525381 External Links: Document Cited by: Ā§1. Khattab and Zaharia (2020) O. Khattab and M. Zaharia ColBERT: efficient and effective passage search via contextualized late interaction over BERT. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, SIGIR 2020, Virtual Event, China, July 25-30, 2020, J. X. Huang, Y. Chang, X. Cheng, J. Kamps, V. Murdock, J. Wen, and Y. Liu (Eds.), p. 39ā48. External Links: Link, Document Cited by: Ā§3. Kim et al. (2024) D. W. Kim, J. S. Park, K. Sharma, A. Velazquez, L. Li, J. W. Ostrominski, T. Tran, R. H. Seitter PerĆ©z, and J. H. Shin Qualitative evaluation of artificial intelligence-generated weight management diet plans. Frontiers in Nutrition 11, p. 1374834. Note: PMID: 38577160; PMCID: PMC10991711 External Links: Document Cited by: Ā§1. Lewis et al. (2020) P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. KĆ¼ttler, M. Lewis, W. Yih, T. RocktƤschel, S. Riedel, and D. Kiela Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin (Eds.), External Links: Link Cited by: Ā§3. Li and Zaki (2020) D. Li and M. J. Zaki RECIPTOR: an effective pretrained model for recipe representation learning. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ā20, New York, NY, USA, p. 1719ā1727. External Links: ISBN 9781450379984, Link, Document Cited by: Ā§1. Mehta et al. (2024) N. H. Mehta, S. L. Huey, R. Kuriyan, J. P. PeƱa-Rosas, J. L. Finkelstein, S. Kashyap, and S. Mehta Potential mechanisms of precision nutrition-based interventions for managing obesity. Advances in Nutrition 15 (3), p. 100186. External Links: ISSN 2161-8313, Document, Link Cited by: Ā§1. Meng et al. (2022) T. Meng, S. Lu, N. Peng, and K. Chang Controllable text generation with neurally-decomposed oracle. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (Eds.), External Links: Link Cited by: Ā§3. Park et al. (2023) J. S. Park, J. C. OāBrien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein Generative agents: interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, UIST 2023, San Francisco, CA, USA, 29 October 2023- 1 November 2023, S. Follmer, J. Han, J. Steimle, and N. H. Riche (Eds.), p. 2:1ā2:22. External Links: Link, Document Cited by: Ā§4. Sahin and Aytekin-Sahin (2024) O. Sahin and G. Aytekin-Sahin Open-source multi-objective optimization software for menu planning. Expert Systems with Applications 252, p. 124213. External Links: ISSN 0957-4174, Document, Link Cited by: Ā§1. Schick et al. (2023) T. Schick, J. Dwivedi-Yu, R. DessƬ, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom Toolformer: language models can teach themselves to use tools. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), External Links: Link Cited by: Ā§3. Tian et al. (2022a) Y. Tian, C. Zhang, Z. Guo, C. Huang, R. A. Metoyer, and N. V. Chawla RecipeRec: A heterogeneous graph learning model for recipe recommendation. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, L. D. Raedt (Ed.), p. 3466ā3472. External Links: Link, Document Cited by: Ā§1. Tian et al. (2022b) Y. Tian, C. Zhang, Z. Guo, Y. Ma, R. A. Metoyer, and N. V. Chawla Recipe2Vec: multi-modal recipe representation learning with graph neural networks. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, L. D. Raedt (Ed.), p. 3473ā3479. External Links: Link, Document Cited by: Ā§1, Ā§3. Wang et al. (2023) X. Wang, J. Wei, D. Schuurmans, Q. V. Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, External Links: Link Cited by: Ā§3. Xie et al. (2026a) M. Xie, Y. Hu, H. Xu, X. Li, Y. Dong, R. Wang, J. Guo, J. He, H. Chen, J. Guo, S. Liu, P. Wang, F. Ren, Z. Fan, J. A. Martinez, and R. Zhu Advances, challenges, and future directions in precision nutrition by machine learning-based prediction of glycemic responses to foods: a narrative review. The American Journal of Clinical Nutrition, p. 101322. External Links: Document, Link Cited by: Ā§1. Xie et al. (2026b) M. Xie, X. Zhang, Y. Li, and C. Lv Structure guided retrieval-augmented generation for factual queries. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), San Diego, California, United States, p. 40350ā40370. External Links: Document, Link Cited by: Ā§3. Yang Yuexin (2018) N. I. f. N. a. H. C. C. Yang Yuexin Chinese food composition tables (standard edition, 6th edition, volume 1). Peking University Medical Press. External Links: ISBN 978-7-5659-1699-1 Cited by: Ā§2, Ā§5. Yang Yuexin (2019) N. I. f. N. a. H. C. C. Yang Yuexin Chinese food composition tables (standard edition, 6th edition, volume 2). Peking University Medical Press. External Links: ISBN 978-7-5659-1978-7 Cited by: Ā§2, Ā§5. Yao et al. (2023) S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. R. Narasimhan, and Y. Cao ReAct: synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023, External Links: Link Cited by: Ā§3. Zhang et al. (2026) X. Zhang, Y. Wan, Y. Li, M. Xie, and C. Lv MC-RAG system: a structure-driven rag system for multi-constraint queries. arXiv preprint arXiv:2607.10151. External Links: 2607.10151, Link Cited by: Ā§3. Zhang et al. (2025) Z. Zhang, Z. Wang, T. Ma, V. S. Taneja, S. Nelson, N. H. L. Le, K. Murugesan, M. Ju, N. V. Chawla, C. Zhang, and Y. Ye MOPI-HFRS: a multi-objective personalized health-aware food recommendation system with llm-enhanced interpretation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 2860ā2871. External Links: Document, Link Cited by: Ā§1.