Pith. sign in

REVIEW 4 major objections 5 minor 42 references

A Novel Approach to Balance Convenience and Nutrition in Meals With Long-Term Group Recommendations and Reasoning on Multimodal Recipes and its Implementation in BEACON

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A boosted bandit algorithm can craft meal plans that match user constraints better than random or sequential baselines.

desk verdict 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. read the letter →

arxiv 2412.17910 v2 pith:EK62RC7E submitted 2024-12-23 cs.LG cs.AI

Vansh Nagpal , Siva Likitha Valluru , Kausik Lakkaraju , Nitin Gupta , Zach Abdulrahman , Andrew Davison , Biplav Srivastava This is my paper · ORCID
classification cs.LGcs.AI
keywords mealrecommendationcontextualbanditsR3reciperepresentationmultimodalrecipesfoodcomputingnutritionLLMtext-to-JSONconversionBEACON
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [BEACON Recommendation Evaluation; Boosted Bandit Algorithm appendix] 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.
  2. [BEACON Recommendation Evaluation, Table 5] 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.
  3. [Fully-automated LLM conversion of recipes to R3 (RC2); Table 3 vs Table 4] 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.
  4. [Goodness Metrics for Recommendations] 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.
minor comments (5)
  1. [R3 Evaluation Metrics] The term 'syntatic similarity score' contains a typo; it should be 'syntactic similarity score.'
  2. [LLM-based Conversion Methods / Evaluation] 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.
  3. [Table 4] 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.
  4. [Discussion and Conclusion] 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.
  5. [Problem Formulation] 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.

Circularity Check

2 steps flagged · score 6.0 of 10

M2's advantage is partly tautological: the user-constraint metric scores the same preference-feature alignment the bandit is trained to predict, so the headline comparison is not an independent test of recommendation quality.

  1. self definitional [Goodness Metrics for Recommendations, User Constraint Metric (uc)]
    "For each meal mi, we calculate a user constraint score uci by comparing the user’s preference with the meal’s ingredient content. If the user’s preference is negative (−1) and the meal contains the ingredient, this counts negatively towards the score. Conversely, if the user’s preference is positive (+1) and the meal contains the ingredient, this counts positively towards the score."

    This metric is not an external quality measure: it scores exactly the alignment of user ternary preferences (dairy/meat/nuts) with item feature flags. The appendix constructs the bandit's training labels from precisely this alignment, so the metric measures the bandit's training target rather than independent meal quality.

  2. self definitional [Boosted Bandit Algorithm for Meal Recommendation (Appendix)]
    "Given a set of users with their dietary preferences(U) as well as the set of all recipes in R3 format(R), we create predicate logic pairs of the form preference(user5, negative nuts) and item(food18, has nuts)... Following this, we also create negative and positive predicate logic pairs of the form recommendation(user18, food22), to signify if a recommendation is a positive or negative recommendation to the user based on their aligning preferences."

    The training signal for M2 is generated from the same preference-feature alignment that uc computes. Since M0 and M1 ignore preferences, M2 is essentially trained to optimize uc, and the reported uc advantage (and the combined uc·dm·mc score) follows by construction. The train/test split avoids exact identity but does not break the definitional correspondence: the concept learned is the concept being evaluated.

full rationale

The central evaluation claim that M2 is 'the most effective in recommending meals' rests on goodness metrics whose user-constraint component is constructed from the same preference-to-ingredient alignment that generates the bandit's training labels. The paper's own appendix says recommendation(user, item) labels are made 'based on their aligning preferences,' and the uc metric rewards exactly that alignment; M0 and M1 do not use preferences at all, so M2's uc advantage is largely forced by the setup. This is partial circularity rather than full identity because a train/test split is used and the bandit must still generalize, and because the paper also contributes an R3 recipe conversion pipeline, a small dataset, and a prototype system. However, the future-work item 'conducting qualitative evaluation to show the acceptance of our recommendation system' confirms that no user-facing or externally grounded validation supports the headline comparison. The R3 conversion evaluation and self-citations to prior R3 and bandit work are not load-bearing circular steps; the circularity is confined to the recommendation-effectiveness comparison.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

Most of the paper's contribution is empirical; it relies on prior representations (R3), a prior algorithm (boosted bandits), and synthetic user data. The main free parameters are the unspecified goodness weights and the unstated bandit hyperparameters.

free parameters (3)
  • Goodness score weights = not reported
    The overall goodness G is a weighted sum of dm, mc, and uc, with 'weights tailored to user preferences'; the actual weights are never specified, making them free parameters.
  • RC2 conversion configuration = 1-shot, e1, temperature 0.3
    Selected post hoc as the only configuration out of 24 trials that returned valid JSON; it is then used as the representative RC2 converter.
  • Bandit hyperparameters = not reported
    The relational boosted bandit implementation details (learning rate, number of estimators, exploration parameters) are not disclosed, so results depend on unstated choices.
assumptions (3)
  • domain assumption R3 is an adequate structured recipe representation for meal recommendation
    The paper builds all data and evaluation on R3 from Pallagani et al. 2022 without validating that this representation captures the nutritional and convenience factors needed for good recommendations.
  • domain assumption Boosted bandits can be applied to meal recommendation by converting user preferences and food features into predicate logic pairs
    The appendix assumes that 'recommendation(user,item)' labels can be derived from preference and item predicates, and that these labels are learnable and predictive of meal quality.
  • ad hoc to paper Simulated user preference distributions reflect realistic user populations
    Configurations c1, c2, c3 use 24 synthetic users with randomly assigned preferences, which is an assumption made by the authors for the evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Novel Approach to Balance Convenience and Nutrition in Meals With Long-Term Group Recommendations and Reasoning on Multimodal Recipes and its Implementation in BEACON." pith.science (2026). https://pith.science/paper/EK62RC7E

@misc{pith2026241217910,
  author       = {Pith},
  title        = {Pith review of: A Novel Approach to Balance Convenience and Nutrition in Meals With Long-Term Group Recommendations and Reasoning on Multimodal Recipes and its Implementation in BEACON},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EK62RC7E}},
  note         = {Machine review of arXiv:2412.17910}
}
read the original abstract

A common decision made by people, whether healthy or with health conditions, is choosing meals like breakfast, lunch, and dinner, comprising combinations of foods for appetizer, main course, side dishes, desserts, and beverages. Often, this decision involves tradeoffs between nutritious choices (e.g., salt and sugar levels, nutrition content) and convenience (e.g., cost and accessibility, cuisine type, food source type). We present a data-driven solution for meal recommendations that considers customizable meal configurations and time horizons. This solution balances user preferences while accounting for food constituents and cooking processes. Our contributions include introducing goodness measures, a recipe conversion method from text to the recently introduced multimodal rich recipe representation (R3) format, learning methods using contextual bandits that show promising preliminary results, and the prototype, usage-inspired, BEACON system.

Figures

Figures reproduced from arXiv: 2412.17910 by the authors.

Figure 1
Figure 1. A brief depiction of solution components used in the BEACON prototype. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Left: Meal preference configuration showing Anthony Gibbson’s selected constraints, including demographic infor [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Performance of meal recommendation methods across configurations, time frames, and metrics. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Representing ingredients in R3 1 { 2 "Calories": { 3 "measure": "417", 4 "unit": "kcal" 5 } 6 } [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Representing nutrition in R3 1 { 2 "original_text": "Beat the egg well. Egg drop soup is noted for its strands of shredded egg . To achieve this characteristic look and texture , make sure you blend the egg mixture well, and stir it slowly into the broth.", 3 "input_co…
Figure 7
Figure 7. Figure 7: An LLM user prompt for extracting instructions in [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: An LLM user prompt for annotating recipes with [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: RC2 system prompt for 1-shot prompting and e2 example type [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: RC2 LLM user prompt for recipe translation from plain-text to R3 BEACON Implementation System Architecture and Design This section describes the architectural design of the web ap￾plication, which employs a hybrid approach combining Re￾act for the frontend and Django …
Figure 11
Figure 11. Figure 11: Calendar-based meal visualization screen on the [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Calendar-based meal visualization screen on the [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: Calendar-based meal visualization screen on the [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 35 canonical work pages

  1. [1]

    Bekdash, R. A. 2024. Epigenetics, Nutrition, and the Brain: Improving Mental Health through Diet. International Journal of Molecular Sciences, 25(7): 4036

    work page 2024

  2. [2]

    Chitale, P.; Gala, J.; and Dabre, R. 2024. An Empirical Study of In-context Learning in LLM s for Machine Translation. In Ku, L.-W.; Martins, A.; and Srikumar, V., eds., Findings of the Association for Computational Linguistics: ACL 2024, 7384--7406. Bangkok, Thailand: Association for Computational Linguistics

    work page 2024

  3. [3]

    Cremonesi, P.; Koren, Y.; and Turrin, R. 2010. Performance of recommender algorithms on top-n recommendation tasks. In Proceedings of the 4th ACM Conf. on Rec. Sys., 39--46

    work page 2010

  4. [4]

    Dong, Q.; Li, L.; Dai, D.; Zheng, C.; Ma, J.; Li, R.; Xia, H.; Xu, J.; Wu, Z.; Liu, T.; et al. 2022. A survey on in-context learning. arXiv preprint arXiv:2301.00234

    arXiv 2022

  5. [5]

    Dunn, C.; Haubenreiser, M.; Johnson, M.; Nordby, K.; Aggarwal, S.; Myer, S.; and Thomas, C. 2018. Mindfulness approaches and weight loss, weight maintenance, and weight regain. Current obesity reports, 7: 37--49

    work page 2018

  6. [6]

    EatThisMuch. 2024. The Automatic Meal Planner - Eat This Much. Accessed: 2024-11-10

    work page 2024

  7. [7]

    Escarda-Fern \'a ndez, M.; L \'o pez-Riob \'o o-Botana, I.; Barro-Tojeiro, S.; Padr \'o n-Cousillas, L.; Gonzalez-V \'a zquez, S.; Carreiro-Alonso, A.; and G \'o mez-Area, P. 2024. LLMs on the Fly: Text-to-JSON for Custom API Calling. Proceedings of the SEPLN-CEDI

    work page 2024

  8. [8]

    Forouzandeh, S.; Rostami, M.; Berahmand, K.; and Sheikhpour, R. 2024. Health-aware food recommendation system with dual attention in heterogeneous graphs. Computers in Biology and Medicine, 169: 107882

    work page 2024

Show all 42 references
  1. [9]

    Ge, M.; Elahi, M.; Ferna \'a ndez-Tob \' as, I.; Ricci, F.; and Massimo, D. 2015. Using tags and latent factors in a food recommender system. In Proceedings of the 5th international conference on digital health 2015, 105--112

    2015

  2. [10]

    Q.; Sablayrolles, A.; Roux, A.; Mensch, A.; Savary, B.; Bamford, C.; Chaplot, D

    Jiang, A. Q.; Sablayrolles, A.; Roux, A.; Mensch, A.; Savary, B.; Bamford, C.; Chaplot, D. S.; de las Casas, D.; Hanna, E. B.; Bressand, F.; Lengyel, G.; Bour, G.; Lample, G.; Lavaud, L. R.; Saulnier, L.; Lachaux, M.-A.; Stock, P.; Subramanian, S.; Yang, S.; Antoniak, S.; Scao...

    2024 arXiv

  3. [11]

    Kakadiya, A.; Natarajan, S.; and Ravindran, B. 2021. Relational boosted bandits. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, 12123--12130

    2021

  4. [12]

    Legislature, C. S. 2018. California Consumer Privacy Act (CCPA). https://oag.ca.gov/privacy/ccpa

    2018

  5. [13]

    Leme, A. C. B.; Hou, S.; Fisberg, R. M.; Fisberg, M.; and Haines, J. 2021. Adherence to Food-Based Dietary Guidelines: A Systemic Review of High-Income and Low- and Middle-Income Countries. Nutrients, 13(3), 1038. https://doi.org/10.3390/nu13031038

    2021 doi

  6. [14]

    Li, D.; and Zaki, M. J. 2020. Reciptor: An effective pretrained model for recipe representation learning. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 1719--1727

    2020

  7. [15]

    Ma, H.; Zhang, C.; Bian, Y.; Liu, L.; Zhang, Z.; Zhao, P.; Zhang, S.; Fu, H.; Hu, Q.; and Wu, B. 2024. Fairness-guided few-shot prompting for large language models. Advances in Neural Information Processing Systems, 36

    2024

  8. [16]

    A.; Kris-Etherton, P

    Metz, J. A.; Kris-Etherton, P. M.; Morris, C. D.; Mustad, V. A.; Stern, J. S.; Oparil, S.; Chait, A.; Haynes, R. B.; Resnick, L. M.; Clark, S.; et al. 1997. Dietary compliance and cardiovascular risk reduction with a prepared meal plan compared with a self-selected diet. The A...

    1997

  9. [17]

    Miaschi, A.; Brunato, D.; Dell ' Orletta, F.; and Venturi, G. 2021. What Makes My Model Perplexed? A Linguistic Investigation on Neural Language Models Perplexity. In Agirre, E.; Apidianaki, M.; and Vuli \'c , I., eds., Proceedings of Deep Learning Inside Out (DeeLIO): The 2nd...

    2021

  10. [18]

    Min, W.; Jiang, S.; and Jain, R. 2019. Food Recommendation: Framework, Existing Solutions and Challenges. IEEE Transactions on Multimedia, PP: 1--1

    2019

  11. [19]

    of Agriculture, U. D. ???? Eat healthy with myplate

  12. [20]

    I.; and (NIST), T

    of Standards, N. I.; and (NIST), T. 2022. Risk Management Framework for Information Systems and Organizations: A System Life Cycle Approach for Security and Privacy. https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-37r2.pdf

    2022

  13. [21]

    N.; and Srivastava, B

    Pallagani, V.; Ramamurthy, P.; Khandelwal, V.; Venkataramanan, R.; Lakkaraju, K.; Aakur, S. N.; and Srivastava, B. 2022. A Rich Recipe Representation as Plan to Support Expressive Multi Modal Queries on Recipe Content and Preparation Process. arXiv preprint arXiv:2203.17109

    2022 arXiv

  14. [22]

    Papastratis, I.; Stergioulas, A.; Konstantinidis, D.; Daras, P.; and Dimitropoulos, K. 2024. Can ChatGPT provide appropriate meal plans for NCD patients? Nutrition, 121: 112291

    2024

  15. [23]

    Pecune, F.; Callebert, L.; and Marsella, S. 2020. A Recommender System for Healthy and Personalized Recipes Recommendations. In HealthRecSys@ RecSys, 15--20

    2020

  16. [24]

    Peeperkorn, M.; Kouwenhoven, T.; Brown, D.; and Jordanous, A. 2024. Is Temperature the Creativity Parameter of Large Language Models? arXiv:2405.00492

    2024 arXiv

  17. [25]

    Ribeiro, D.; Machado, J.; Ribeiro, J.; Vasconcelos, M. J. M.; Vieira, E. F.; and De Barros, A. C. 2017. SousChef: mobile meal recommender system for older adults. In International Conference on Information and Communication Technologies for Ageing Well and e-Health, volume 2, ...

    2017

  18. [26]

    Rostami, A. 2024. An Integrated Framework for Contextual Personalized LLM-Based Food Recommendation. Ph.D. thesis, UC Irvine

    2024

  19. [27]

    T.; Valluru, S

    Srivastava, B.; Koppel, T.; Paladi, S. T.; Valluru, S. L.; Sharma, R.; and Bond, O. 2022. ULTRA: A Data-driven Approach for Recommending Team Formation in Response to Proposal Calls. In 2022 IEEE International Conference on Data Mining Workshops (ICDMW), 1002--1009. IEEE

    2022

  20. [28]

    Srivastava, B.; and Pallagani, V. 2024. The Case for Developing a Foundation Model for Planning-like Tasks from Scratch. arXiv:2404.04540

    2024 arXiv

  21. [29]

    Su, X.; and Khoshgoftaar, T. M. 2009. A survey of collaborative filtering techniques. Advances in AI, 2009

    2009

  22. [30]

    Union, E. 2016. General Data Protection Regulation (GDPR). https://gdpr-info.eu/

    2016

  23. [31]

    L.; Srivastava, B.; Paladi, S

    Valluru, S. L.; Srivastava, B.; Paladi, S. T.; Yan, S.; and Natarajan, S. 2024 a . Promoting Research Collaboration with Open Data Driven Team Recommendation in Response to Call for Proposals. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 22833--22841

    2024

  24. [32]

    L.; Widener, M.; Srivastava, B.; and Gangopadhyay, S

    Valluru, S. L.; Widener, M.; Srivastava, B.; and Gangopadhyay, S. 2024 b . ULTRA: Exploring Team Recommendations in Two Geographies Using Open Data in Response to Call for Proposals. In Proceedings of the 7th Joint International Conference on Data Science & Management of Data ...

    2024

  25. [33]

    L.; Widener, M.; Srivastava, B.; Natarajan, S.; and Gangopadhyay, S

    Valluru, S. L.; Widener, M.; Srivastava, B.; Natarajan, S.; and Gangopadhyay, S. 2024 c . AI-assisted research collaboration with open data for fair and effective response to call for proposals. AI Magazine

    2024

  26. [34]

    Yagcioglu, S.; Erdem, A.; Erdem, E.; and Ikizler-Cinbis, N. 2018. R ecipe QA : A Challenge Dataset for Multimodal Comprehension of Cooking Recipes. In Riloff, E.; Chiang, D.; Hockenmaier, J.; and Tsujii, J., eds., Proceedings of the 2018 Conference on Empirical Methods in Natu...

    2018

  27. [35]

    P.; Dell, N.; Belongie, S.; Cole, C.; and Estrin, D

    Yang, L.; Hsieh, C.-K.; Yang, H.; Pollak, J. P.; Dell, N.; Belongie, S.; Cole, C.; and Estrin, D. 2017. Yum-me: a personalized nutrient-based meal recommender system. ACM Transactions on Information Systems (TOIS), 36(1): 1--31

    2017

  28. [36]

    Zeevi, D.; Korem, T.; Zmora, N.; Israeli, D.; Rothschild, D.; Weinberger, A.; Ben-Yacov, O.; Lador, D.; Avnit-Sagi, T.; Lotan-Pompan, M.; et al. 2015. Personalized nutrition by prediction of glycemic responses. Cell, 163(5): 1079--1094

    2015

  29. [37]

    Zhang, H.; Dong, Y.; Xiao, C.; and Oyamada, M. 2024. Large Language Models as Data Preprocessors. arXiv:2308.16361

    2024 arXiv

  30. [38]

    Q.; and Artzi, Y

    Zhang, T.; Kishore, V.; Wu, F.; Weinberger, K. Q.; and Artzi, Y. 2020. BERTScore: Evaluating Text Generation with BERT. arXiv:1904.09675

    2020 arXiv

  31. [39]

    Zhang, X.; Rajabi, N.; Duh, K.; and Koehn, P. 2023. Machine translation with large language models: Prompting, few-shot learning, and fine-tuning with qlora. In Proceedings of the Eighth Conference on Machine Translation, 468--481

    2023

  32. [40]

    Zioutos, K.; Kondylakis, H.; and Stefanidis, K. 2023. Healthy Personalized Recipe Recommendations for Weekly Meal Planning. Computers, 13(1): 1

    2023

  33. [41]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  34. [42]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.