{"id":"05ef3255-a75b-423c-a523-cf4c835903e9","arxiv_id":"2412.17910","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"In simulated evaluations, a boosted bandit meal recommender outperforms random and sequential baselines on the authors' new goodness metrics.","lead":"This paper introduces BEACON, a meal recommendation system that uses a contextual bandit algorithm to balance nutrition and convenience. The authors evaluate it on a small set of converted fast-food and soul-food recipes, scoring meal plans with new goodness metrics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"M2's advantage is partly tautological: the user-constraint metric scores the same preference-feature matches the bandit is trained to predict, so the evaluation does not independently validate meal quality.","rationale":"The reader's weakest assumption identified the synthetic-user setup and the choice of metrics as load-bearing. My read agrees that the evaluation is the weak point, but sharpens the issue: the uc metric is not merely one possible measure of meal quality, it is essentially the same objective the bandit was trained to optimize. The bandit learns to predict whether a user will prefer an item based on feature-preference matches; uc rewards exactly those matches. Therefore M2's superiority on uc and on the combined metric is expected and does not provide evidence that M2 produces better meal plans in any user-relevant sense. The baselines are deliberately preference-blind, so the comparison is between a model optimized for the metric and models that ignore the metric. This is a form of evaluation leakage rather than a claim that is simply outside current consensus. However, the paper's contributions extend beyond the M2 comparison: the R3 conversion study (RC0/RC1/RC2) is evaluated with independent semantic/syntactic metrics, and the BEACON prototype is a concrete system. The conditional verdict is still appropriate, but the conditions should include decoupling the bandit's training features from the evaluation features and ideally adding a human or held-out user evaluation. I would not reject the paper outright, but the 'most effective' claim needs to be softened or the evaluation needs to be made non-circular. Thus I recommend keeping the reader's CONDITIONAL verdict.","tokens_in":16985,"tokens_out":3321,"duration_ms":33834,"concrete_test":"Hold out one of the three food features entirely from bandit training (e.g., train the bandit on dairy and meat preferences only, excluding nuts), then compute uc using all three features including nuts. If M2's combined-score advantage persists when evaluated on an unseen feature, the result is not purely an artifact of training/evaluation alignment; if the advantage shrinks or reverses, the current conclusion is an artifact. Ideally, also compare M2 against a preference-aware baseline that uses the same features but a different algorithm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the boosted bandit (M2) is 'the most effective in recommending meals' rests on a comparison where the primary metric is aligned with M2's training objective. In the 'User Constraint Metric (uc)' section, uc is computed by comparing user ternary preferences on hasDairy, hasMeat, and hasNuts with each meal's ingredient flags. The appendix 'Boosted Bandit Algorithm for Meal Recommendation' states that the bandit is trained on predicate pairs such as preference(user, negative nuts) and item(food, has nuts), with recommendation(user, food) labels derived from aligning preferences with item features. Thus the training signal and the uc evaluation metric are constructed from the same preference-feature alignment. M0 and M1 ignore preferences entirely, so M2's higher uc, and consequently its higher combined uc·dm·mc score, is largely by construction. The paper's own future-work item 4 ('conducting qualitative evaluation to show the acceptance of our recommendation system') confirms that no user-facing validation was performed. This does not invalidate the R3 conversion contribution or the system prototype, but it means the headline comparison is not an externally meaningful test of recommendation effectiveness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BEACON, a meal-planning system that balances convenience and nutrition through customizable meal configurations and time horizons. It introduces a standardized multimodal recipe representation (R3), converts 52 recipes from text to R3 using manual, hybrid-LLM, and fully automated methods, and evaluates three recommendation algorithms (random M0, sequential M1, and a relational boosted bandit M2) on simulated users across three preference configurations and three time horizons. The main claimed contribution is that the boosted bandit is the most effective meal recommender, supported by user-constraint, duplicate, and meal-coverage metrics, plus a prototype web application.","tokens_in":17221,"tokens_out":3281,"duration_ms":32545,"significance":"The problem of long-horizon, configurable meal recommendation is relevant, and the R3 conversion experiments with LLMs are of practical interest: the paper's observation that freely available LLMs struggle to produce valid R3 JSON, and that hybrid methods with manual curation perform better, is a useful, falsifiable data point. The BEACON prototype and the release of a 52-recipe R3 dataset are concrete artifacts. However, the central comparative claim about the boosted bandit's effectiveness rests entirely on a simulated evaluation whose primary metric (uc) is constructed from the same preference-feature alignment the bandit is trained to predict, so the headline result is not an independent validation of recommendation quality. If the evaluation were strengthened with held-out user feedback, independent metrics, and statistical rigor, the system-level contribution would be solid.","major_comments":[{"comment":"The user-constraint metric uc is computed by comparing user ternary preferences on hasDairy, hasMeat, and hasNuts with meal ingredient flags, while the boosted bandit is trained on predicate pairs of exactly that form (e.g., preference(user, negative nuts) and item(food, has nuts)), as described in the appendix. Consequently, M2's higher uc scores, and therefore its higher combined metrics in Table 5, are partly by construction, and the conclusion that 'the boosted bandit algorithm represented by M2 is the most effective in recommending meals' (Evaluation section) is not independently supported. The paper should add an evaluation signal not derived from the same preference-feature pairs, such as a small user study, expert dietitian assessment, or a holdout set of real user choices.","section":"BEACON Recommendation Evaluation; Boosted Bandit Algorithm appendix"},{"comment":"Table 5 reports only point estimates for each metric, with no variance, error bars, or significance tests, despite the experiments involving 24 simulated users per configuration and three configurations. Differences such as c3, t1 bandit uc = 0.992 versus random uc = 0.963 may be within sampling noise, so the claim that M2 is the most effective method is not statistically substantiated. The authors should report means and standard deviations over multiple random seeds or user samples, together with significance tests or confidence intervals.","section":"BEACON Recommendation Evaluation, Table 5"},{"comment":"The representative RC2 converter (1-shot, e1, temperature 0.3) was selected by evaluating 24 configurations on 5 recipes (Table 3), but the final comparison in Table 4 evaluates that converter on the 52 original recipe texts. If the 5 tuning recipes are included in the 52, the RC2 metrics in Table 4 are optimistically biased. The paper must state whether the 5 recipes used for configuration selection are disjoint from the 52-recipe evaluation set; if they are not, a held-out split must be used for the conversion comparison.","section":"Fully-automated LLM conversion of recipes to R3 (RC2); Table 3 vs Table 4"},{"comment":"The goodness score G is defined as a weighted sum of md, cs (later uc?), and uc with 'weights tailored to user preferences,' but the evaluation in Table 5 reports values like 'uc dm mc', 'uc dm', 'uc mc', and 'dm mc' without ever specifying the formulas or the weights. This ambiguity makes the combined metrics impossible to interpret; the authors should define the exact aggregation (e.g., whether these are products, averages, or weighted sums) and justify the weight choices.","section":"Goodness Metrics for Recommendations"}],"minor_comments":[{"comment":"The term 'syntatic similarity score' contains a typo; it should be 'syntactic similarity score.'","section":"R3 Evaluation Metrics"},{"comment":"The text says 'we evaluate the performance ... of 5 recipes across these 24 configurations and present them in Figure 3,' but the conversion results are actually shown in Table 3; the cross-reference should be corrected.","section":"LLM-based Conversion Methods / Evaluation"},{"comment":"In Table 4, the RC2 entry labeled 'RC2α26' appears to have a data-formatting artifact; the value 0.376 likely belongs to the jec column, but the table should be cleaned for readability.","section":"Table 4"},{"comment":"The conclusion states the authors 'displayed a dataset of 50 R3 items,' but earlier the paper says 52 recipes in R3 format; the count should be made consistent.","section":"Discussion and Conclusion"},{"comment":"The notation for user preferences mixes 'likesDairy' in the problem formulation with 'hasDairy' as the feature in the evaluation; please standardize these terms to avoid confusion.","section":"Problem Formulation"}],"recommendation":"major_revision","confidential_remarks":"The paper is best framed as a systems and data-contribution paper; the recommendation-algorithm comparison is not yet at journal strength. The central concern is construct validity of the uc metric rather than a simple lack of baselines, and it should be addressed directly. The R3 conversion study and the BEACON prototype are within scope and provide a reasonable basis for revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real systems paper, not a fake one, but the central empirical claim needs to be walked back or re-supported. The problem formulation (variable meal configurations, flexible time horizons), the three goodness metrics, and the BEACON prototype are legitimate contributions. The most solid piece is the recipe conversion study: comparing manual, hybrid-LLM, and fully automated LLM conversion of recipes into R3, and finding that only 1 of 24 Mixtral configurations produced valid JSON while hybrid ChatGPT with manual assembly did better, is a concrete and useful result. That alone justifies a look.\n\nThe soft spot is the bandit evaluation. The uc metric compares user ternary preferences on dairy/meat/nuts to meal ingredient flags; the appendix shows the bandit is trained on exactly those preference-feature pairs to produce recommendation scores. So M2's high uc is largely by construction, and the conclusion that M2 is \"most effective\" is overstated. The other metrics partially undercut it: M2 scores worst on duplicates, and its advantage in the combined score is driven by uc. There are also no error bars or significance tests, only 24 synthetic users per configuration with randomly assigned preferences, and the paper itself defers qualitative user evaluation to future work. These do not invalidate the system or the conversion study, but they mean the recommendation comparison is not an independent test of meal quality.\n\nA minor issue: the RC2 configuration was selected on just 5 recipes and then applied to the full 52; that is not fatal but should be acknowledged. Also, no code or data is released, which limits reproducibility.\n\nWho is this for? People building meal-planning tools, working on recipe representation, or wanting a worked example of an LLM-to-structured-JSON pipeline. For a RecSys or applied ML venue, it is a borderline accept with major revisions; for a systems or demo track, it is reasonably solid. I would send it to serious peer review, with the expectation that the authors either strengthen the evaluation or soften the claim about M2 being most effective.","headline":"A genuine systems paper with a useful problem formulation and a solid LLM-to-R3 conversion study, but its headline bandit comparison is built on an evaluation metric that tracks the bandit's own training signal.","tokens_in":17761,"tokens_out":2211,"would_cite":false,"duration_ms":23882,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A boosted bandit algorithm can craft meal plans that match user constraints better than random or sequential baselines.","keywords":["meal recommendation","contextual bandits","R3 recipe representation","multimodal recipes","food computing","nutrition","LLM text-to-JSON conversion","BEACON"],"falsifier":"Run the same three methods with real users or a richer preference model on a held-out set of meal plans and measure the goodness scores; if the boosted bandit no longer beats random or sequential selection on user-constraint and meal-coverage metrics, the paper's central claim fails. A simpler check: if users rate M2's plans no higher than M0's in a blind preference test, the advantage is an artifact of the scoring function.","tokens_in":16805,"feed_emoji":"🍽️","tokens_out":5735,"duration_ms":50748,"temperature":0.7,"pith_summary":"Meal choice is usually treated as a tradeoff between nutrition and convenience, and most people do not stick to dietary guidelines. This paper argues that a data-driven meal recommender can balance those pressures by letting users specify meal configurations, ingredient preferences, and time horizons, then generating multi-day plans from a structured recipe representation. The paper's central evidence is a simulation in which a boosted-bandit method, called M2, outscores random and sequential baselines on user-constraint and meal-coverage metrics across one-, three-, and five-day plans. It also reports that freely available LLMs alone cannot reliably convert plain-text recipes into the structured R3 JSON format, so semi-automated or rule-checked conversion is needed. The authors position this as making long-term meal recommendation a solvable planning problem rather than a one-off food suggestion task.","feed_headline":"Boosted bandit beats random and sequential meal planners","feed_subtitle":"A preference-aware algorithm keeps meal plans on target across 1-, 3-, and 5-day horizons.","key_machinery":"The load-bearing object is the R3 (Rich Recipe Representation) format, a structured JSON schema for recipes that captures ingredients, nutrition, food roles, and atomic cooking-instruction tasks with input conditions, output quality, tools, and failure states, along with image and video modalities. On top of it, the recommendation machinery is a relational boosted bandit (M2) that learns predicate-style preference and item-relation scores, such as preference(user5, negative nuts) and item(food18, has nuts), and converts them into recommendation probabilities to assemble meals. The third piece is the goodness metric, a weighted combination of duplicate score (diversity within and across meals), meal coverage score (requested food roles fulfilled), and user-constraint score (ingredient content matches like, neutral, or dislike flags). Together these make the problem measurable and let the bandit be compared with random and sequential baselines.","core_discovery":"The paper's central claim is that long-term meal recommendation should be treated as a group recommendation problem over structured, multimodal recipes, and that a relational boosted bandit is the best of three tested methods for producing such plans. In the authors' evaluation with 24 synthetic users per configuration, M2 outperforms random selection (M0) and sequential rotation (M1) on the user-constraint and meal-coverage components of their goodness score, while M1 alone always scores perfectly on avoiding duplicate items. The authors also conclude that LLM-based text-to-R3 conversion is unreliable in its fully automated form: only one of twenty-four configurations of Mixtral-8x7B-Instruct-v0.1 produced valid JSON, and manually curated recipes score highest on semantic preservation. The BEACON prototype is presented as a usage-inspired implementation that lets a user configure preferences and view, edit, and track meal plans on a calendar.","pith_inferences":["Extending beyond the paper, the same bandit-plus-R3 pipeline could be tested on real user feedback, where the ternary like, neutral, and dislike model would likely need expansion to allergens, macronutrients, and budget constraints.","The R3 representation is a planning artifact, so the approach may transfer to adjacent tasks such as grocery shopping lists, restaurant menu filtering, or institutional meal planning for schools and hospitals.","One testable extension is to add a diversity-aware exploration bonus to M2 and check whether the combined goodness score can beat M2 while keeping its user-constraint advantage.","The evaluation's synthetic users make the headline result a proof-of-concept; a real-user study with measured adherence would be the natural next validation and would reveal whether the goodness metrics track what people actually value."],"forward_implications":["If M2's advantage is real, meal planners can be built that adapt to user preferences with no hand-written dietary rules, only preference flags and structured recipe data.","The R3 format makes it possible to reason simultaneously about what a dish contains and how it is prepared, enabling queries such as diabetic-friendly, culturally relevant meals that are easy to cook.","Because LLM-only conversion fails on complex JSON, practical recipe ingestion must pair LLM extraction with validation, repair, or human curation, which bounds the cost of scaling the recipe dataset.","The bandit's lower duplicate score indicates that satisfying preferences can lead to repeated items, so a production system must decide how much repetition users tolerate.","Longer horizons up to five days do not erase M2's advantage, supporting the idea that the method generalizes to weekly meal planning."],"supporting_citations":[{"why":"Introduces the R3 rich recipe representation that the paper adopts and extends for all recipe data and reasoning.","marker":"Pallagani et al. 2022"},{"why":"Supplies the relational boosted bandits algorithm that M2 adapts to meal recommendation.","marker":"Kakadiya, Natarajan, and Ravindran 2021"},{"why":"Describes SousChef, the closest prior meal recommender, which the paper compares against on configurability and horizon.","marker":"Ribeiro et al. 2017"},{"why":"Provides the meta-survey showing widespread non-adherence to dietary guidelines, motivating the problem.","marker":"Leme et al. 2021"},{"why":"Supports the premise that following a provided meal plan reduces health risks compared with self-selected diets.","marker":"Metz et al. 1997"},{"why":"Supplies the RecipeQA dataset from which the original twenty-five R3 recipes are derived.","marker":"Yagcioglu et al. 2018"},{"why":"Defines BERTScore, used as the semantic similarity metric for evaluating R3 conversion quality.","marker":"Zhang et al. 2020"}],"fun_headline_variants":["Bandit beats random and sequential for long-term meal plans","Boosted bandit outperforms simple planners in meal group recommendations","BEACON's bandit method wins on balanced meal planning over baselines","Meal recommender: boosted bandit tops random and sequential baselines","Data-driven meal planning: bandit beats rotation and random selection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that 24 simulated users per configuration, with randomly assigned like, neutral, or dislike for dairy, meat, and nuts, are a valid stand-in for real users, and that the duplicate, coverage, and user-constraint scores capture what matters in a meal plan.","fun_headline_variants_meta":{"raw":{"variants":["Bandit beats random and sequential for long-term meal plans","Boosted bandit outperforms simple planners in meal group recommendations","BEACON's bandit method wins on balanced meal planning over baselines","Meal recommender: boosted bandit tops random and sequential baselines","Data-driven meal planning: bandit beats rotation and random selection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1377,"prompt_tokens":882,"completion_tokens":495,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":405}},"tokens_in":498,"tokens_out":495,"duration_ms":5424,"temperature":1.0,"reasoning_tokens":405,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:08:05.324758+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same three methods with real users or a richer preference model on a held-out set of meal plans and measure the goodness scores; if the boosted bandit no longer beats random or sequential selection on user-constraint and meal-coverage metrics, the paper's central claim fails. A simpler check: if users rate M2's plans no higher than M0's in a blind preference test, the advantage is an artifact of the scoring function.","supporting_citations":[{"cited_title":"A Rich Recipe Representation as Plan to Support Expressive Multi Modal Queries on Recipe Content and Preparation Process","cited_arxiv_id":"2203.17109","evidence_quote":"Introduces the R3 rich recipe representation that the paper adopts and extends for all recipe data and reasoning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the relational boosted bandits algorithm that M2 adapts to meal recommendation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes SousChef, the closest prior meal recommender, which the paper compares against on configurability and horizon."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the meta-survey showing widespread non-adherence to dietary guidelines, motivating the problem."},{"cited_title":"A.; Kris-Etherton, P","cited_arxiv_id":null,"evidence_quote":"Supports the premise that following a provided meal plan reduces health risks compared with self-selected diets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the RecipeQA dataset from which the original twenty-five R3 recipes are derived."}],"review_version":1}