Source-linked AI summary
Profile-Aware Trustworthy Recipe Generation with Planner-Critic Agentic Remediation
Shanhong Liu, Pai Chet Ng, Konstantinos N. Plataniotis
TL;DR
Food-image recipe generation often overlooks whether a recipe meets a particular user's allergy, dietary, and preparation-safety constraints. PCAR separates recipe planning from safety auditing and uses structured feedback to revise unsafe drafts. Results show strong safe-generation performance and practical quality with capable backbones, alongside reliability differences across models.
Problem
Traditional one-pass recipe generation does not explicitly enforce user-profile constraints or verify safety, while adding profile information to prompts may still miss constraints or hidden risks.
Source: §1 · p. 1Method
PCAR uses a Planner Agent to generate image- and profile-conditioned drafts and a Safety Critic Agent to audit them and send structured feedback for remediation.
Source: §1 · pp. 1–2Results
GPT-4o achieves 98% safe generation success and 2% unsafe leakage, while results indicate that reliability depends strongly on the backbone model.
Source: §3.3 · pp. 4–5Takeaways & Limitations
PCAR can achieve high safe-generation success while preserving practical recipe quality when supported by capable backbone models.
Source: §4 · p. 5Performance remains highly dependent on the underlying model, and lightweight local models show weaker stability.
Source: §4 · p. 5
Abstract
from arXiv · showhide
Recipe generation from food images has practical value for intelligent cooking assistance, but traditional one-pass generation often overlooks user-specific safety requirements such as allergies, dietary restrictions, and preparation constraints. We propose PCAR, a Planner-Critic Agentic Remediation framework for trustworthy recipe generation. PCAR separates recipe planning from safety verification: a Planner Agent extracts ingredients and generates recipe drafts conditioned on the user profile, while a Safety Critic Agent audits each draft and provides structured feedback for remediation when violations are detected. This remediation loop enables unsafe recipes to be revised rather than directly returned or discarded. We evaluate PCAR on real food images with 100 benchmark user profiles across four backbone models, including proprietary and locally served open-source models. Results show that PCAR achieves strong safety and generation performance with capable backbone models, while preserving practical recipe quality.
1. INTRODUCTION
Food-image recipe systems can recognize ingredients and generate recipes, yet personalized cooking assistance also needs to account for whether a recipe is safe for a specific user. PCAR frames recipe generation as balancing practical utility with profile-aware safety and uses planner–critic remediation to revise unsafe drafts.
1. INTRODUCTION
- Existing food-image research has advanced recognition, ingredient inference, and recipe generation, but has paid less attention to safety for specific users.The cited work spans recognition, recommendation, dietary tracking, retrieval, and generation.Source: §1 · p. 1
- PCAR treats recipe generation as a constrained task that balances practical recipe utility with user-specific safety requirements.The constraints include allergies, dietary restrictions, and preparation safety.Source: §1 · pp. 1–2
- A Planner Agent generates image- and profile-conditioned drafts, while an independent Safety Critic Agent audits them and returns structured feedback for revision.The loop continues until approval or a bounded remediation limit is reached.Source: §1 · p. 2
- The evaluation covers four backbone models and 100 benchmark user profiles, including baseline constraints, complex combinations, derivative risks, and high-risk scenarios.The models include proprietary and locally served open-source systems.Source: §1 · p. 2
2. PROPOSED METHOD
The method formulates recipe generation as maximizing practical utility within a user-specific safe recipe space. PCAR separates planning from safety auditing and feeds structured critiques back into a bounded revision loop.
2.1. Problem Formulation
- The formulation seeks a practically useful recipe within the safe set for user u, with safety defined over allergens, dietary restrictions, and health or preparation constraints.Utility captures image grounding, culinary coherence, instruction completeness, and personalization; the profile is Pu = (Au, Du, Hu).Source: §2.1 · p. 2
2.2. Planner–Critic Agentic Remediation (PCAR)
- The Planner Agent extracts perceived ingredients from the image and generates drafts conditioned on those ingredients, the user profile, and prior safety feedback.The initial draft uses no prior feedback; later drafts use feedback-conditioned replanning.Source: §2.2 · p. 2
- The Safety Critic Agent audits each draft and returns a binary decision plus structured feedback about detected violations.The violation set covers allergen, dietary, and preparation-safety constraints.Source: §2.2 · pp. 2–3
- Unsafe drafts are iteratively revised toward the safe recipe space, and a maximum remediation budget T bounds the loop.If the budget is exhausted without approval, the policy terminates fail-safe rather than returning an unsafe recipe.Source: §2.2 · p. 3
3. EXPERIMENTS AND EVALUATION
The experiments evaluate PCAR using real food images, 100 varied user profiles, and four backbones. Results vary substantially by backbone, while approved recipes can receive high quality scores.
3.1. Experimental Setup
- The test set pairs randomly sampled Epicurious food images with profiles from four cohorts, covering baseline, combined, derivative-risk, and high-risk constraints.The benchmark comprises 100 profiles, 25 per cohort; cleaned ingredient lists are used to evaluate ingredient perception, not as generation targets.Source: §3.1 · p. 3
3.2. Evaluation protocol
- Four backbones are evaluated in three trials over 100 profiles each, for 300 cases per model, spanning proprietary APIs and locally served open-source models.The local-model setup reflects a consumer-grade or edge-accessible deployment scenario.Source: §3.2 · p. 3, Table 1 · p. 4
3.3. Results and Discussion
- GPT-4o achieves 98% safe generation success with 2% unsafe leakage and no crashes; Gemini-2.5-Flash reaches 74% SGSR, while local models score lower.Table 1 reports results averaged over 300 cases per model; Gemini's unsafe leakage and crash rates are 11.33% and 14.67%, respectively.Source: Table 1 · p. 4, §3.3 · pp. 4–5
3.2. Evaluation protocol
- Evaluation measures semantic ingredient F1, approved safe-generation success and allergen leakage, remediation counts, recipe quality, and end-to-end latency.Quality is judged on six dimensions for successful approved recipes; failed, crashed, or unsafe cases are excluded.Source: §3.2 · p. 4
3.3. Results and Discussion
- GPT-4o achieves 100% SGSR with no unsafe leakage in Baseline Control and High-Risk Scenarios, and remains strong in the other cohorts.Local models show lower cohort-level robustness, especially Qwen2.5-VL-3B on Complex Combinations and High-Risk Scenarios.Source: Table 2 · p. 5, §3.3 · p. 5
- Gemini-2.5-Flash has the highest overall recipe quality score at 4.5357, while approved recipes are generally practical and coherent.Quality judgments cover successful approved recipes only; Gemma3-4B-IT's high average rests on three judged recipes, making its estimate unstable.Source: Table 3 · p. 5, §3.3 · p. 5
- GPT-4o combines the highest safe-generation success with the lowest latency; Qwen2.5-VL-3B is faster than Gemma3-4B-IT but much less safe.The reported comparison frames backbone choice as a safety–efficiency trade-off.Source: Figure 3 · p. 5, §3.3 · p. 5
4. CONCLUSION
Trustworthy recipe generation must satisfy user-specific dietary, allergy, and preparation-safety constraints, not merely produce visually plausible recipes. PCAR achieves high safe-generation success while preserving practical recipe quality, though performance depends on the backbone model.
4. CONCLUSION
- PCAR achieves high safe-generation success while preserving practical recipe quality.The conclusion reports this outcome without specifying numerical results.Source: §4 · p. 5
- Trustworthy recipe generation must account for user-specific dietary, allergy, and preparation-safety constraints beyond visual plausibility.Source: §4 · p. 5
- Performance remains highly dependent on the underlying model, and lightweight local models show weaker stability.Future work targets local-model reliability, richer nutrition and medical constraint modeling, and improved ingredient perception.Source: §4 · p. 5