{"id":"9de66c0c-c87f-4952-b03e-03acfff00631","arxiv_id":"1909.00105","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A sequence-to-sequence model that conditions on a user's past recipes generates more personalized and plausible cooking instructions from a recipe name and partial ingredients.","lead":"This paper introduces a new machine-learning task: generating complete cooking instructions from just a recipe name, a few ingredients, and a user's past recipe reviews. It shows that conditioning on user history improves the plausibility and personalization of generated recipes over a non-personalized baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Personalization claim rests on likelihood-ranking proxy; human judges were not target users, so 'plausible and personalized' conflates two distinct properties.","rationale":"The reader flagged the reliability of review histories as the weakest premise; I agree that is a risk, but the more load-bearing gap is downstream: even granting that histories encode preferences, the paper never measures whether conditioning on them produces recipes the target user prefers. UMA is an intrinsic likelihood-ranking metric that can be satisfied by a model that has merely learned to encode user identity in its output distribution; it does not measure the value of that conditioning to the user. The anonymous pairwise evaluation validates plausibility but cannot validate personalization, because judges are not given the user context that the model was conditioned on. Thus the abstract's 'plausible and personalized' claim is not jointly supported by the reported experiments. This is an evaluation-validity concern rather than an internal inconsistency, and it is addressable with a targeted user study, so the correct response is to keep the reader's CONDITIONAL stance rather than accept or reject outright. The model, dataset, and released code are real contributions; the condition is that the personalization half of the central claim needs direct user-side evidence.","tokens_in":8965,"tokens_out":6360,"duration_ms":81771,"concrete_test":"Run a target-user preference study on the held-out split: for each sampled test user, generate two recipes from the same name, ingredient, and calorie input—one conditioned on the user's own historical reviews and one conditioned on a control user's history of similar length—present the pair in randomized order to the target user, and ask which recipe they would prefer to cook or follow. If self-conditioned recipes are not chosen at a rate significantly above 50% with a confidence interval across users, the personalization claim is unsupported. To make this feasible, the study can be limited to 100–150 users and the control profile can be the non-personalized baseline output; the decisive comparison is target-user preference, not anonymous-judge coherence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Personalization is the load-bearing half of the central claim, and it is supported only by an indirect likelihood proxy. The pairwise human evaluation (Table 2, 310 pairs per model) asks anonymous judges to choose the more coherent recipe or the one that best accomplishes the recipe name; those judges are not the target users and are not shown the target user's history, so a 63% preference rate demonstrates general plausibility, not personalized fit. The UMA/MRR metric in Section 5 computes the likelihood of a generated recipe under the gold user versus nine randomly generated profiles. This only shows the generator's output distribution is sensitive to conditioning on user histories; it does not show the target user would prefer or even recognize the generated recipe, and it cannot rule out that the model succeeds by exploiting shallow profile cues such as frequent ingredients or techniques while producing generic instructions. The reported gains also lack error bars and significance tests, and the Neural Checklist baseline numbers are omitted after being called 'comparable.' The missing step is a direct test that conditioning on a user's history changes output in a way the target user actually values.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the task of personalized recipe generation: given a recipe name, a few ingredients, a calorie level, and a user's historical recipe interactions, the model generates full recipe instructions tailored to that user. The proposed architecture encodes the recipe name and ingredients with BiGRUs, applies ingredient attention during decoding, and incorporates user information either by attending over the k most recent prior recipes, prior recipe names, or a technique-preference vector; an attention fusion layer combines these contexts before the final output softmax. The authors release a Food.com dataset of 180K recipes and 700K interactions and evaluate against a nearest-neighbor baseline and an encoder-decoder baseline, reporting perplexity, BLEU/ROUGE, distinctness, user matching accuracy (UMA), mean reciprocal rank (MRR), automatic coherence and entailment scores, and pairwise human preference. They report that personalized models beat the baselines on perplexity, diversity, UMA/MRR, coherence, and a 63% human pairwise preference rate.","tokens_in":9214,"tokens_out":5299,"duration_ms":91493,"significance":"If the claims hold, this is a useful new task connecting natural language generation and recommender systems, with a novel dataset and a clean decomposition of user modeling into recipe-level, name-level, and technique-level signals. The public release of code and data is a strength, as are the automatic coherence metrics and the UMA/MRR personalization probes. However, the paper's central claim has two parts, plausibility and personalization, and the evidence for the second half is indirect: the human judges were not the target users, and UMA measures likelihood ranking under different conditioning profiles rather than actual user preference. The absence of error bars and significance tests further weakens the quantitative comparisons. With additional validation, the task and model would be a solid contribution.","major_comments":[{"comment":"The UMA/MRR result demonstrates that the model's output distribution changes with the conditioning user profile, but it does not establish that the generated recipe is more appropriate for the target user than for another user. The pairwise human evaluation uses anonymous judges who see only the recipe name and ingredients, not the target user's history, so it tests plausibility, not personalized fit. Add a target-user study (for example, each user chooses between the personalized and baseline recipe for their own held-out review) or validate UMA against a behavioral signal such as the user's subsequent review or rating of the generated recipe.","section":"Section 5, Personalization"},{"comment":"All quantitative results are reported as point estimates without error bars, confidence intervals, or significance tests, and several margins are small (for example, BPE PPL 9.516 vs. 9.551 for Prior Name vs. Enc-Dec, and coherence 1.82 vs. 1.77). Report multiple random seeds or bootstrapped intervals and paired significance tests, or explicitly state that the differences are not statistically tested. This is essential for the comparative claims made throughout the paper.","section":"Section 5, Tables 2 and 4"},{"comment":"The paper treats users' Food.com reviews as evidence of recipes they 'consumed,' but a review is not necessarily a record of cooking or liking a recipe. The personalization signal and the UMA evaluation both depend on this assumption, and the paper provides no analysis of review text or ratings to validate it. Please justify or soften this assumption and discuss how review behavior might differ from consumption behavior.","section":"Section 4, Dataset and task framing"},{"comment":"The description of the nine 'randomly generated user profiles' used in UMA evaluation is underspecified. If these profiles are drawn from a different activity distribution (for example, shorter or less typical histories), the gold user could rank highest for reasons unrelated to personalization. Describe the sampling procedure and, ideally, include a control condition with random histories matched on length and recipe popularity.","section":"Section 5, Personalization"}],"minor_comments":[{"comment":"The Neural Checklist model is mentioned as adapted and then dropped because Enc-Dec has 'comparable performance,' but no Neural Checklist numbers are shown; either report them in Table 2 or move them to the appendix for transparency.","section":"Section 5, Baselines"},{"comment":"The manually constructed list of 58 techniques and the string-match detection method are described with no precision about matching rules or false positives; a sentence on how inflected forms and multiword expressions are handled would help.","section":"Section 4, Techniques"},{"comment":"Calorie-level conditioning is part of the model but is never ablated or analyzed; a small experiment or a statement on its effect would clarify whether it contributes to the reported gains.","section":"Section 5"},{"comment":"The NN row reports a dash for BPE PPL with no explanation; please add a note on why that value is omitted.","section":"Table 2"},{"comment":"The recipe-level coherence metric is said to have a maximum score of 2, but the reported scores (1.77-1.82) are all close to each other and to the maximum; consider reporting the distribution or the score of gold recipes as a reference point.","section":"Section 5, Recipe Level Coherence"}],"recommendation":"major_revision","confidential_remarks":"I see no reason to doubt the authors' good faith. The main gap is that the central 'personalized' claim is supported by a likelihood proxy and by anonymous judges rather than by the target users themselves; I believe this is fixable with an additional study or a validation of UMA against a behavioral signal. If such evidence cannot be provided, the claim should be weakened accordingly. The omitted Neural Checklist numbers are a transparency issue, not a sign of misconduct. The paper fits the scope of the journal and would be a solid contribution after the requested revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid new-task paper with a reusable dataset and a mostly honest evaluation, but the personalization claim is softer than the abstract suggests. The UMA/MRR metric is a likelihood-ranking proxy: it shows the model's output distribution shifts when conditioned on different user histories, not that the target user would prefer or recognize the generated recipe. The human evaluation confirms plausibility, not personalized fit. The stress-test note is right on that point.\n\nWhat's actually new: the task formulation — generating recipes from a name, a few ingredients, and a calorie level while conditioning on user history — plus the dataset of 180K recipes and 700K interactions. The model is a standard encoder-decoder with attention, and the Prior Name variant is a simple, sensible way to encode user history without learning item embeddings. That's fine for a task paper. The sequential split and BPE preprocessing are careful, and the qualitative examples show real differences across conditions.\n\nSoft spots, in order of softness. First, the UMA metric, as noted. It's a reasonable discriminative test, but it could be inflated by popularity or generic recipe styles; a simple frequency-based baseline (e.g., always favor a user's most common ingredients or techniques) would be a useful sanity check and is missing. Second, no error bars or significance tests anywhere. The reported gains are modest — a few tenths of perplexity, a few points of coherence — and could easily be noise, especially with only 310 pairwise comparisons per model. Third, the Neural Checklist baseline is adapted and then dropped with \"comparable performance\" but no numbers; readers can't verify that claim. Fourth, the automatic coherence and entailment metrics are used without validation against human judgments; they might be measuring fluency rather than step order. These are all surmountable.\n\nThe central task definition holds up, and the paper is refreshingly clear that BLEU/ROUGE are not appropriate for generation quality. The claim \"plausible and personalized\" should be read as \"plausible, and conditioned on user history in a way that is detectable via likelihood.\" That is still a useful contribution.\n\nWho this is for: people working on data-to-text with user conditioning, and the food-computing community. It deserves a serious referee. The main revisions would be adding variance estimates, a frequency-based personalization baseline, and ideally a small study where target users judge recipes generated for them.\n\nRecommendation: send to peer review. I'd accept after moderate revision.","headline":"A useful new-task paper with a reusable dataset, but the personalization claim is softer than the abstract suggests: the UMA metric is a likelihood-ranking proxy, and the human evaluation shows plausibility, not personalized fit.","tokens_in":9687,"tokens_out":1554,"would_cite":false,"duration_ms":17617,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Conditioning recipe generation on a user's past reviews produces recipes judged more plausible and more personal than generation without user history.","keywords":["personalized recipe generation","data-to-text generation","user preference modeling","attention fusion","encoder-decoder generation","recipe coherence","user matching accuracy","incomplete recipe specifications"],"falsifier":"Run a blind comparison in which the same dish is generated once from the user's true review history and once from the history of ten other users with similar culinary tastes, then ask the user which recipe better fits their preference; if the true-history recipe is not chosen significantly above chance, the apparent personalization is mostly genre or cuisine matching.","tokens_in":8807,"feed_emoji":"🍳","tokens_out":10773,"duration_ms":94429,"temperature":0.7,"pith_summary":"The paper proposes personalized recipe generation: turning a dish name, a few key ingredients, and a calorie level into full cooking instructions adapted to a specific user. Its central claim is that attending to a user's history of reviewed recipes is enough to make the output more plausible, more diverse, and more aligned with that user than a generic encoder-decoder recipe generator. To support this claim, the paper contributes a new dataset of over 180K recipes and 700K user reviews, automatic metrics for recipe coherence and personalization, and a pairwise human evaluation in which the personalized outputs were preferred over the non-personalized baseline 63% of the time.","feed_headline":"Tuned to past reviews, recipe generator wins 63% of comparisons","feed_subtitle":"A dish name and a few ingredients become a cooking guide matched to the cook's own review history.","key_machinery":"The central mechanism is the attention fusion layer, which at each decoding step concatenates the ingredient context, the prior-recipe or prior-technique context, the decoder hidden state, and the previous token embedding, then applies a ReLU projection and softmax to produce the next-token distribution. Prior-recipe attention averages over the user's k=20 most recent reviewed recipes, represented either by recipe id or by the mean of the recipe name's token embeddings, while prior-technique attention adds a user-technique co-occurrence prior to the attention score. This is what injects user history into the generation process and what distinguishes the three personalized variants: Prior Recipe, Prior Name, and Prior Tech.","core_discovery":"The paper's central discovery is that conditioning recipe generation on a user's historical recipe interactions improves the plausibility, diversity, and personalization of the generated instructions relative to a non-personalized encoder-decoder. A user is represented by the set of recipes they have reviewed, by the names of those recipes, or by the cooking techniques occurring in them; at each decoding step the model attends over the user's k most recent reviewed recipes and fuses that context with ingredient attention and decoder state. In user-matching tests, the variant attending over prior recipe names ranks the correct user first in about half the cases and best in mean reciprocal rank, and human evaluators preferred the personalized outputs to the non-personalized baseline in 63% of pairwise comparisons.","pith_inferences":["If review history is a reliable taste signal, the same attention-fusion idea should transfer to other content-to-text settings where the \"user\" is defined by feedback, such as shopping lists, workout plans, or medication instructions.","The user-matching experiment could be sharpened by testing users whose histories overlap in cuisine: if a model conditioned on a different but cuisine-similar user scores almost as well, the personalization is largely genre matching rather than per-person taste.","A simple ablation anyone can run is to pass the k recent recipes in shuffled or reversed order; if performance holds, the model is using the set of prior recipes rather than their temporal drift, which would simplify the memory mechanism.","The coherence metrics introduced here, step-order scoring and step entailment, could be applied to other instructional prose, since they do not depend on recipe-specific knowledge."],"forward_implications":["Users with as little as four prior reviews can receive a plausible recipe from a dish name and a few ingredients; the model works even though half of the users have six or fewer reviews.","The most effective personalization signal is the names of previously reviewed recipes, not the recipes' ids or their cooking techniques.","Personalized generation increases recipe diversity and explicit ingredient mentions without sacrificing coherence, as measured by step-order scoring and step entailment.","The dataset and evaluation protocol, including sequential leave-one-out splits, user-matching accuracy, and coherence scorers, can serve as a benchmark for incomplete-specification recipe generation.","Human preference for personalized recipes, 63% over baseline, indicates that user history can steer instructional text generation beyond simple template filling."],"supporting_citations":[{"why":"Supplies the recipe-generation setting and checklist-attention idea this work extends to the personalized and incomplete-input case.","marker":"Kiddon et al. (2016)"},{"why":"Provides the additive attention scoring used for ingredient and prior-recipe contexts.","marker":"Bahdanau et al. (2015)"},{"why":"Contributes the pairwise human evaluation and the user-ranking likelihood methodology used to measure personalization.","marker":"Fan et al. (2018)"},{"why":"Supplies the neural recipe-coherence scoring model used to evaluate step ordering of generated recipes.","marker":"Bosselut et al. (2018a)"},{"why":"Provides the collection of 384 cooking actions from which the 58 techniques in the technique-attention model were manually derived.","marker":"Bosselut et al. (2018b)"},{"why":"BERT encodes recipe steps in the coherence scorer and the step-entailment classifier.","marker":"Devlin et al. (2019)"},{"why":"Gives the sequential leave-one-out evaluation protocol that keeps test interactions out of the training set.","marker":"Kang and McAuley (2018)"}],"fun_headline_variants":["Recipe AI tailors cooking steps to your review history","Your past recipe picks shape new personalized instructions","Recipe generator that learns from your taste, wins 63% of tests","Historical preferences guide recipe generation for better personal fit","AI that reads your food history to craft custom cooking guides"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that what a person reviewed in the past reliably predicts what they want in a new dish, even when their review history contains as few as four items.","fun_headline_variants_meta":{"raw":{"variants":["Recipe AI tailors cooking steps to your review history","Your past recipe picks shape new personalized instructions","Recipe generator that learns from your taste, wins 63% of tests","Historical preferences guide recipe generation for better personal fit","AI that reads your food history to craft custom cooking guides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000138,"raw_usage":{"total_tokens":1076,"prompt_tokens":787,"completion_tokens":289,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":403,"completion_tokens_details":{"reasoning_tokens":211}},"tokens_in":403,"tokens_out":289,"duration_ms":21136,"temperature":1.0,"reasoning_tokens":211,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T06:01:20.368900+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a blind comparison in which the same dish is generated once from the user's true review history and once from the history of ten other users with similar culinary tastes, then ask the user which recipe better fits their preference; if the true-history recipe is not chosen significantly above chance, the apparent personalization is mostly genre or cuisine matching.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the recipe-generation setting and checklist-attention idea this work extends to the personalized and incomplete-input case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the pairwise human evaluation and the user-ranking likelihood methodology used to measure personalization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BERT encodes recipe steps in the coherence scorer and the step-entailment classifier."}],"review_version":1}