cs.AIOct 29, 2026

Can Personal Requirements Become Supervision for Planning?

Authors: Junwei Wu, Runze Yan, Yuzhang Xie, Hanqi Luo, Ziwei Dong, Jiaying Lu, Xiao Hu, Carl Yang

Organizations: Emory University, Atlanta, Georgia, USA · Microsoft, San Francisco, California, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Managing multiple chronic conditions through diet requires satisfying disease- and medication-specific requirements jointly across meals and days while preserving established eating habits. These requirements are public in clinical guidelines and drug labels, yet turning them into a feasible plan for one person remains difficult for language-model agents. In our dietitian-approved weekly planning evaluation built from NHANES, thirteen models, from medically fine-tuned 8B models to GPT-6.1 Sol, pass fewer than 3% of complete plans through direct generation. Explicit rules, higher reasoning effort, or model-directed tools leave GPT-6.1 Sol at no more than 8%. We present Noura, which pairs LLM planning with executable personal specifications instantiated from public clinical guidance and individual health and dietary profiles. During planning, the specification grounds foods, optimizes portions, checks the complete multi-day plan, and guides revision. Our key insight is that a specification that rejects a plan can also determine what to learn from it. Noura turns executed plans and contextually credited revisions into supervision for compact models, recovering targets even from trajectories that never achieve complete success. The same specification supplies supervised targets, preferences, and rewards for post-training. On 4,993 held-out participants, Noura-SFT raises Qwen3-14B's complete-plan pass rate from 24.0% to 72.5% within the same execution loop, cuts rule violations by 97.5% with no observed critical failures, and improves dietary quality. The locally hosted 14B planner incurs fewer rule violations than its GPT-6 Luna teacher. A matched comparison shows that executed targets reduce rule violations by 41% over model-written proposals. Across four Qwen, Phi, and Llama backbones, self-training alone reduces rule violations by 41-61% on a 1,249-person population subset.

Figures & tables

 Personal requirementsExecutionLearning
MethodProfileGuidancePreferencesHorizonCheckSolveReviseSignalSelfTarget
Maillot et al. ( Maillot et al., 2010 )✗✓✓✗✓✓✗–––
Amiri et al. ( Amiri et al., 2023 )✓✓✓✗✓✓✗–––
NutriOrion ( Wu et al., 2026 )✓✓✓✗✓✗✗–––
AMIE ( Liévin et al., 2026 )✓✓✗✗✗✗✗S, P✗✗
LLM+P ( Liu et al., 2023 )✗✗✗✗✓✓✗–––
Table 1. Comparison of personal requirements, execution, and learning (Section 2 ). S/P/R denote supervised/preference/reward signals; Target indicates learning the executed plan.
Figure 1. Noura connects execution to learning: (1) execute and revise plans under a personal specification; (2) construct targets, preferences, and rewards; (3) update the planning policy. Solid arrows show computation or data flow, the dashed return shows LoRA updates, and the circled plus denotes prompt concatenation. Construction versions and checkpoint lineage appear in Appendix A . A three-stage diagram shows planning under a personal specification, construction of executed targets, preferences, and rewards from trajectories, and LoRA policy updates that feed back into planning.
ConfigurationCPR ↑\uparrowDPR ↑\uparrowRV ↓\downarrowCVP ↓\downarrowHEI ↑\uparrow
Qwen3-14B
Direct generation0.518.640,0768.0–
Base + execution24.090.01,5870.080.0
Noura-SFT (self)41.095.36410.077.4
Noura-DPO (self)50.695.95470.075.2
Noura-SFT72.599.1390.083.4
Table 2. Population evaluation on 4,993 held-out participants. CPR, DPR, and CVP are percentages; RV counts failed rule-day checks; HEI is mean HEI-2020. Bold marks the best Qwen3-14B values; – denotes unevaluated HEI.
TargetRowsUpdatesCPR ↑\uparrowRV ↓\downarrowDPR ↑\uparrowHEI ↑\uparrow
A. Self-generated: target content
Model-written3,63145628.321385.576.8
Executed3,64145630.712590.478.1
B. Teacher-generated: full construction
Initial drafts1,50018821.313576.782.9
Executed + revisions3,56344668.71498.581.4
Table 3. Supervision controls for Qwen3-14B on 300 test participants. A shares trajectories, admission with complete revision credit, and 3,631 prompts. B uses the same GPT-6 Luna trajectories from 1,500 training participants. Rows and updates report filtered training examples and optimizer steps, respectively. CPR/DPR are percentages.
Figure 2. Supervision beyond complete success. Training participants are grouped by admitted target type, from up to 1,500 trajectories per model (Phi-4: 1,499). Marks indicate the success-only pool, and multipliers compare full-plan target coverage with that pool. Corpus construction is detailed in Appendix A . Stacked bars show training participants contributing passing full-plan targets, improved full-plan targets, revision targets only, or no target across five source models.
Figure 3. Planning profiles on 300 participants. Cells show absolute values (k: thousands); shading shows within-column normalized changes from each backbone (blue: favorable; red: unfavorable). Metric definitions appear in Appendix B.3 . Numerical profiles of 21 planning configurations show ten metrics for outcomes, revisions, generation cost, and dietary quality. Shading compares each trained policy with its backbone. Reference and medical models are unshaded.
Figure 4. Planning quality versus output tokens per participant on the same 300 participants. Lines link each base to its self-trained checkpoint; token comparisons are within backbone. Thinking and other trained variants use Qwen3-14B; API points provide quality references. Two scatter panels plot complete-plan passing and rule violations against output tokens per participant across local backbones, with API-model quality references.
Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

FamilyApplies whenImplemented constraintEvidence basis
SodiumHypertension; heart failure≤2,300{\leq}2{,}300 ; ≤2,000{\leq}2{,}000 mg/dayAHA/ACC hypertension ( Jones et al., 2025 ) ; AHA/ACC/HFSA heart failure ( Heidenreich et al., 2022 )
KidneyCKD stage G3 or higherPhosphorus ≤1,000{\leq}1{,}000 mg/day; protein ≤0.8{\leq}0.8 g/kg/dayKDIGO CKD-MBD ( Kidney Disease: Improving Global Outcomes CKD-MBD Update Work Group(2017), KDIGO ) ; KDIGO 2024 ( Kidney Disease: Improving Global Outcomes CKD Work Group(2024), KDIGO )
Added sugarDiabetes≤10%{\leq}10\% of energyADA Standards of Care ( American Diabetes Association Professional Practice Committee for Diabetes, 2026 )
Potassium, upper †eGFR <30{<}30 ; serum K ≥5.0{\geq}5.0 mmol/L; potassium-sparing diuretics; potassium chloride≤2,000{\leq}2{,}000 mg/day (kidney, serum K); ≤2,400{\leq}2{,}400 mg/day (drugs)KDIGO 2024 ( Kidney Disease: Improving Global Outcomes CKD Work Group(2024), KDIGO ) ; drug labels
Potassium, lowerLoop or thiazide diuretic with serum K <3.5{<}3.5 mmol/L≥3,400{\geq}3{,}400 (men), ≥2,600{\geq}2{,}600 (women) mg/day; upper bounds prevailDrug labels; NASEM ( National Academies of Sciences, Engineering, and Medicine, 2019 )
GrapefruitSimvastatin, lovastatin; nifedipineNo grapefruit, pomelo, or Seville orange foodsDrug labels
Appendix
Table 4. Scored requirements by family. Sources provide the clinical basis; numerical bounds include study operationalizations. Drug labels are FDA structured product labels from DailyMed ( National Library of Medicine, nd ) ; † critical severity.
 Dir.RulesToolsExec.
ModelCPRRVCPRRVCPRRVCPRRV
Qwen3-8B0.01,937––––16.7124
Qwen3-14B0.01,9600.01,9380.01,69318.7121
Qwen3-32B0.01,638––––38.785
Phi-40.01,470––0.0 ‡1,354 ‡48.050
gpt-oss-20b0.01,148––0.01,14737.3107
Appendix
Table 5. Planning conditions on 150 participants: CPR (%) / RV (count). Dir.: direct generation; Rules: rules in prompt; Tools: model-directed tools; Exec.: specification-guided execution; –: not evaluated. ‡ 4,000-token output limit (6/150 replies reached it).
ConfigurationCPR ↑\uparrowRV ↓\downarrowCVP ↓\downarrowX-day ↓\downarrowHEI ↑\uparrowABT ↓\downarrow
Direct generation0.03,80026.71,31773.94.2
Rules + self-review0.73,78327.71,511–20.3
Check feedback3.01,82114.3–76.0–
Initial-plan optimization1.07051.32,17680.17.1
Complete execution loop17.02320.067479.338.3
w/o portion optimization4.71,37211.01,12476.939.4
Appendix
Table 6. Qwen3-14B execution controls on 300 participants. CPR and CVP in %; X-day: cross-day findings; ABT: seconds per participant; –: not reported.
 CPR ↑\uparrowRV ↓\downarrowΔ\Delta RV [95% CI]DPR ↑\uparrowHEITok. ↓\downarrow
Qwen3-14B23.7352 90.579.927.5
self-SFT40.7143−209  [−282,−140]-209\;[-282,-140]95.877.416.3
Phi-459.7111 95.571.518.4
self-SFT71.465−46  [−77,−16]-46\;[-77,-16]98.071.013.1
Qwen3-8B27.5569 92.259.529.2
self-SFT28.4223−346  [−456,−240]-346\;[-456,-240]95.259.519.4
Appendix
Table 7. Base and self-SFT on the 1,249-person subset (all complete, none critical). Δ\Delta RV: paired [95% CI]. Tokens: k per participant, within family.
ModelMedQAMedMCQAPubMedQAMMLU-MMMLU-GIFEval
14B Noura-SFT (self)+0.2+0.2−0.5-0.5+1.2+1.2−0.2-0.2−1.0-1.0−5.7-5.7
14B Noura-DPO (self)+0.2+0.2−0.6-0.6+0.2+0.20.00.0−1.1-1.1−5.2-5.2
14B Noura-SFT+0.1+0.1−0.8-0.8+0.6+0.6−0.4-0.4−0.9-0.9−3.3-3.3
Phi-4−0.3-0.3+0.4+0.40.00.0−0.2-0.2+0.3+0.3+3.9+3.9
Qwen3-8B+0.4+0.4−1.1-1.1−0.6-0.6+0.1+0.1+0.2+0.2−1.8-1.8
Llama-3.1-8B+2.0+2.0+3.3+3.3+1.0+1.0+2.6+2.6+1.9+1.9+4.4+4.4
Appendix
Table 8. Change from base (points): position-debiased QA, strict IFEval; MMLU-M/G: medical/general. Other rows: self-SFT (Llama: earlier checkpoint).
More in this topic

LLM Planning

Large language model planning studies how models decompose goals, choose actions, and revise plans under constraints.

See all 147 papers

Explore similar work

CardsList

  1. ShanLiangRen: A Nutrition Agent for Personalized Daily Meal Planning

    Oct 6, 2026Miao Xie, Xiao Zhang, Yuan Wang +2Multi-Objective OptimizationLLM Agent Planning

    College of Information and Electrical Engineering, China Agricultural University · Key Laboratory of Precision Nutrition and Food Quality, China Agricultural University

  2. NutriOrion: A Hierarchical Multi-Agent Framework for Personalized Nutrition Intervention Grounded in Clinical Guidelines

    Feb 20, 2026Junwei Wu, Runze Yan, Hanqi Luo +7Multi-Agent LLM SystemsClinical Decision Support

    Emory University, Atlanta, GA, USA · Texas Woman’s University, Denton, TX, USA · Cecelia Health, New York, NY, USA

  3. Investigating the Ability of Large Language Models to Analyze Recipes for Diabetes

    Sep 3, 2026Revathy Venkataramanan, Aditya Luthra, Venkatesan Nadimuthu +1LLM EvaluationHealthcare

    AI Institute, University of South Carolina 1112 Greene Street, Columbia, SC 29201 USA