acceptodds
Under review as a conference paper at ICLR 2027

NutriAgent-RL: Learning to Route Rule-grounded Agent Workflows for Personalized Clinical Nutrition Guidance

Abstract

Personalized clinical-nutrition guidance should adapt general dietary knowledge to each patient’s clinical context. For an AI system providing such guidance, the context determines whether it should provide dietary advice directly, add a caution, or recommend referral. We use clinical-nutrition redline rules to describe how patient conditions and risk triggers change the response to a dietary query. Each rule specifies which advice should be avoided and what response the system should provide. Applying such rules requires determining which evidence to retrieve and which checks to perform for each query. A common design applies the same large language model (LLM) agent workflow to every query, even when cases differ in their need for retrieval and checking. This design wastes computation on simple cases and leaves relevant constraints unchecked in more complex ones. Thus, workflow selection is central to redline-aware generation. NutriAgent-RL addresses this limitation by learning a query-dependent policy over complete agent workflows. Its contextual-bandit router uses parsed patient context to choose among four complete workflows: Direct Answer, retrieval-augmented generation (RAG), Multi-agent Reasoning, and Fixed Revision. During offline training, rule-grounded critics use applicable rule evidence to score the completed outputs of candidate workflows. These scores, together with penalties for over-refusal and inference cost, provide route-level rewards for offline policy learning of the router. To evaluate the learning of workflow selection, we further construct ND-Bench with 45,033 tasks generated from 449 canonical rules grounded in public nutrition resources and guideline-derived evidence. Compared with RAG, the strongest fixed workflow on Qwen2.5-7B-Instruct, NutriAgent-RL lowers the benchmark-defined escape rate by 7.1%, while using 49.8% fewer total tokens per task.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Personalized clinical nutrition guidance, Reinforcement learning, Agent workflows routing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kCVLOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper develops the NutriAgent-RL framework to formulate redline-aware nutrition guidance generation as a route-level workflow selection problem. Specifically, given patient context and redline evidence, a contextual bandit router learns to select from among 4 workflows - direct answer, RAG, multi agent reasoning and fixed revision. The goal is to select the optimal route that meets redline rule constraints while also optimizing computation time. The paper creates a benchmark of tasks and clinician nutrition redline rules and tests the framework using the benchmark compared to baselines.

Critical strengths
  • The paper is framed well, with justification for the need of the approach (i.e., optimized routing of personalized nutrition advice), and novelty as compared to previous works.
  • The methodology appears sound and the experimental results appear well thought out and provide evidence the method works better than all baselines.
  • The paper has very strong presentation, with a clearly motivated problem, and helpful graphics.
Critical weaknesses
  • My only concern is related to K, the redline rule bank. Where do these rules come from? How were they validated to ensure they were sufficient in both validity (i.e., they cover appropriate clinical requirements) and breadth (i.e., they cover a sufficient breadth in the types of rules relevant for all nutrition guidance cases.)?
Questions for authors
  • Where do the redline rules come from? How were they validated to ensure they were sufficient in both validity (i.e., they cover appropriate clinical requirements) and breadth (i.e., they cover a sufficient breadth in the types of rules relevant for all nutrition guidance cases.)?
  • How much patient context is needed for the methodology to work? Does the method degrade if the patient has large amounts of missing contextual data? Section 3.1 makes a comment about how missing or ambiguous information is preserved, but the paper would be strengthened if the exact effect of amount of missing patient context was discussed or quantified.
Reviewer_Po39OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents NutriAgent-RL, a framework that learns to select different agent workflows for personalized clinical nutrition guidance. It uses a contextual-bandit router to choose among direct answering, RAG, multi-agent reasoning, and response revision based on patient context. The router is trained using rule-based rewards that consider safety, response quality, and computational cost. The authors also introduce ND-Bench, a benchmark containing 45,033 tasks based on 449 clinical nutrition rules. The authors evaluate the framework using three 7B language models, comparing it against four fixed workflows in terms of rule compliance, over-refusal, and inference cost. They also conduct ablation studies and evaluate generalization to unseen rules and data sources.

Critical strengths
  1. Adaptive workflow selection: The paper addresses the limitations of fixed agent workflows by dynamically selecting different reasoning strategies based on patient context, which can improve both efficiency and response quality.

  2. Practical problem setting: The paper focuses on personalized clinical nutrition guidance, where patient-specific constraints make safety and rule compliance particularly important.

  3. Computational efficiency: The proposed method significantly reduces token usage compared to fixed RAG workflows while achieving better rule compliance, making adaptive routing more practical.

Critical weaknesses
  1. Lack of adaptive routing baselines: The experiments only compare against four fixed workflows. Without comparisons to heuristic-based or other learned routing methods, it is difficult to determine whether the improvements come from the proposed RL approach or simply from using adaptive workflow selection.

  2. Unclear correctness of workflow selection: The paper evaluates routing accuracy using RouteMatch, which measures agreement with a predefined heuristic. However, it is unclear whether this heuristic reflects the most appropriate workflow for each query, particularly when multiple workflows could produce equally valid responses.

  3. Limited evidence of clinical safety: The evaluation primarily measures compliance with predefined nutrition rules. However, following these rules does not necessarily guarantee clinically safe recommendations, especially when patient information is incomplete or important risks are not covered by the rule bank.

  4. Potential reward exploitation: Although the reward function penalizes over-refusal and computational cost, the paper does not sufficiently investigate whether the router can exploit weaknesses in the critic-based rewards. High reward scores may not always correspond to genuinely safe or useful responses.

  5. Insufficient analysis of routing failures: The paper reports aggregate performance but provides limited analysis of cases where the router selects an inappropriate workflow. Such errors are particularly important in clinical settings, where choosing an insufficiently rigorous workflow may lead to unsafe recommendations.

Questions for authors

Could the authors clarify how they validate that the learned routing policy selects appropriate workflows, beyond agreement with the predefined RouteMatch heuristic, and whether it outperforms simpler heuristic-based routing strategies?

Reviewer_p26WOpenReview ↗
Rating
1
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
1
Read review
Paper summary

This paper focuses on a practical task to provide personalized clinical-nutrition guidance, and proposes a RL-based workflow routing method for different user needs. To evaluate the outcome of the proposed routing method, this paper also create an ND-Bench from public nutrition resources and guideline-derived evidence. Experiments are conducted across three frozen 7B LLMs, and the workflow selection method outperforms the four individual methods on their proposed benchmark. This routing method is claimed to consume less tokens compared to RAG.

Critical strengths
  • The benchmark design is reasonable. It links patient context, applicable rules, and expected response behavior. Including low-risk control cases also helps distinguish rule compliance from unnecessary refusal.
Critical weaknesses
  • Unresolved inconsistency in the main results. For Qwen model, the four candidate DCA scores sum to only 32.9%, yet the router achieves 82.9%. Under matched outputs and judgments, selection alone cannot explain this gap, and the evaluation protocol does not adequately support it. Besides, this might also lead to questioning the authenticity of this paper.

  • Limited novelty. The routing, fixed critic-based reward design, and benchmark form a closely coupled engineering pipeline, which does not support the claimed three individual contribution. The paper niether sufficiently establish a technical advance over simpler adaptive routers nor the benchmark’s independent scientific value.

Questions for authors
  1. Does the router merely choose one of the four baseline answers, or does something else change when that workflow is selected? Since the router only selects among four fixed workflows, how does it achieve 82.9% DCA when their individual accuracies sum to 32.9%? Are the workflows and evaluation settings identical in the fixed and routed experiments? Please clarify differences in generation, judging, or aggregation and provide per-item evidence.

  2. Please clarify whether Multi-agent Reasoning uses retrieved evidence and specify the actual number of LLM calls for each workflow. How were these calls accounted for in the reported costs?

  3. How can we trust the scoring? Which model judges the answers, and do its judgments agree with qualified human assessment?

  4. Do the main findings generalize? Since both holdout splits omit DietContra and Control, what evidence supports generalization of the reported accuracy–over-refusal trade-off?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.