Elamur Research⌕
☰
home / papers / 2026 / Cohere Labs Community at FoodBench-QA 2026: The Cake Makes the Ingredients

Cohere Labs Community at FoodBench-QA 2026: The Cake Makes the Ingredients

Paper · 2026 · Ravi Ranjan, Roshan Santhosh and Lucien Carroll

Summary

The authors address a risk in LLM-based nutrition advice: generative models can invent nutrient values, which may mislead non-expert users. For the FoodBench-QA 2026 shared task, they reframe nutrient estimation and entity linking as retrieval rather than generation. For nutrient and FSA traffic-light prediction, recipes are encoded with TF-IDF over unigrams and bigrams. The system retrieves the 20 nearest training recipes by cosine similarity and combines their values through voting weighted by squared similarity. If any neighbour exceeds 0.95 similarity, a "Golden Match" rule copies that recipe's values directly. Cohere Command R+ then paraphrases the retrieved information.

For NER/NEL, they use a dictionary built from the training data, mapped to FoodOn, SNOMED CT and Hansard identifiers. Matching uses longest-match-first word-boundary regexes plus morphological normalisation. The system was weaker than the leaderboard leader rbls-lab on numeric nutrient accuracy. However, it beat rbls-lab on the hard "amber" safety labels and on CafeteriaFCD entity linking. It struggled on scientific abstracts and synthetic data. The authors conclude that retrieval only helps when the reference data closely matches the target domain.

Key points

  • Retrieval pipeline: TF-IDF (sklearn, 1–2-grams) feeds k-NN retrieval with k=20. Neighbours vote with weight = sim². A Golden Match threshold of 0.95 overrides the vote. Code is published at github.com/rsk2327/FoodBenchQA.
  • Task 1 (nutrient accuracy, ingredients only): The system scored lower than rbls-lab, for example protein 0.7635 vs 0.9344 and fat 0.6745 vs 0.8446. Adding recipe titles changed results only slightly.
  • Task 2.1 (amber FSA labels, F1, ours vs rbls-lab):
    • fat-amber: 0.7499 vs 0.7086
    • salt-amber: 0.7434 vs 0.6570
    • saturates-amber: 0.7072 vs 0.5767
    • Including titles slightly lowered the system's scores, e.g. salt-amber fell to 0.7302.
  • Task 3 (CafeteriaFCD, macro-F1, ours vs rbls-lab):
    • Hansard: 0.7786 vs 0.7525
    • FoodOn: 0.7859 vs 0.7742
    • SNOMED: 0.8008 vs 0.7848
    • Recall was poor on scientific abstracts and on the artificial dataset.
  • Ablations and limitations:
    • The morphology engine raised NER recall by 18%.
    • Squared weights beat linear weights.
    • Boosting high-calorie keywords did not help.
    • Known weaknesses are insensitivity to portion size, mishandling of negation (e.g. "no added sugar"), and dependence on how well the knowledge base covers the dishes.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🏢Adobe SystemsSan Jose · United States35 works · 3,711 citations🏢CohereSan Francisco · United States15 works · 1,223 citations

Also on the byline: GLA University.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksInformation Systems · 1,437 works

🧭 Topics 1Research subjects and languages the work is about.Question answering · 441 works

Cite

Cite as Ravi Ranjan, Roshan Santhosh and Lucien Carroll. 2026. Cohere Labs Community at FoodBench-QA 2026: The Cake Makes the Ingredients. Proceedings of the Third Workshop on Patient-Oriented Language Processing (CL4Health) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4mxctoq6yb8s

@inproceedings{ranjan2026cohere,
  title = {{Cohere Labs Community at FoodBench-QA 2026: The Cake Makes the Ingredients}},
  author = {Ravi Ranjan, Roshan Santhosh and Lucien Carroll},
  booktitle = {Proceedings of the Third Workshop on Patient-Oriented Language Processing (CL4Health) @ LREC 2026},
  year = {2026},
  doi = {10.63317/4mxctoq6yb8s},
  url = {https://doi.org/10.63317/4mxctoq6yb8s}
}

References

14 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Marc Alexander, Mark Davies, and Fraser Dallachy. 2015. The Hansard Corpus 1803-2005. find
  2. Mohamed Ali, Zaki Taha, and Mohamed Mabrouk Morsey. 2026. Ontology-grounded knowledge graphs for mitigating hallucinations in large language models for clinical question answering. Journal of Biomedical Informatics, page 104993. find
  3. Kevin Donnelly. 2006. SNOMED-CT: The advanced terminology and coding system for eHealth. Studies in Health Technology and Informatics, 121:279–290. find
  4. Damion Dooley, Liliana Andrés-Hernández, Georgeta Bordea, Leigh Carmody, Duccio Cavalieri, Lauren Chan, Pol Castellano-Escuder, Carl Lachat, Fleur Mougin, Francesco Vitali, Chen Yang, Magalie Weber, Hande Kucuk McGinty, and Matthew Lange. 2024. OBO Foundry food ontology interconnectivity. Semantic W… find
  5. Damion M. Dooley, Emma J. Griffiths, Gurinder S. Gosal, Pier L. Buttigieg, Robert Hoehndorf, Matthew C. Lange, Lynn M. Schriml, Fiona S. L. Brinkman, and William W. L. Hsiao. 2018. FoodOn: a harmonized food ontology to increase global food traceability, quality control and data integration. npj Scie… find
  6. Ana Gjorgjevikj, Matej Martinc, Gjorgjina Cenikj, Sašo Džeroski, Barbara Koroušić Seljak, and Tome Eftimov. 2025. FoodSEM: Large Language Model Specialized in Food Named-Entity Linking. In Discovery Science, pages 395–410, Cham. Springer Nature Switzerland. find
  7. Gordana Ispirova, Gjorgjina Cenikj, Matevž Ogrinc, Eva Valenčič, Riste Stojanov, Peter Korošec, Ermanno Cavalli, Barbara Koroušić Seljak, and Tome Eftimov. 2022. CafeteriaFCD Corpus: Food Consumption Data Annotated with Regard to Different Food Semantic Resources. Foods, 11(17):2684. tion, From a Pe… find
  8. Language Resource References Tome Eftimov and Ana Gjorgjevikj and Matej Martinc and Gjorgjina Cenikj and Sašo Džeroski and Barbara Koroušić Seljak. 2026. FoodBench-QA: Shared Task on Grounded Food & Nutrition Question Answering - Codabench. CodaBench. find
  9. Yubin Kim, Hyewon Jeong, Shan Chen, Shuyue Stella Li, Chanwoo Park, Mingyu Lu, Kumail Alhamoud, Jimin Mun, Cristina Grau, Minseok Jung, Rodrigo Gameiro, Lizhou Fan, Eugene Park, Tristan Lin, Joonsik Yoon, Wonjin Yoon, Maarten Sap, Yulia Tsvetkov, Paul Liang, Xuhai Xu, Xin Liu, Chunjong Park, Hyeonho… — [arXiv:2503.05777](https://arxiv.org/abs/2503.05777) find
  10. OpenAI. 2026. OpenAI: AI as a Healthcare Ally. Technical report, Institution. find
  11. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:282… find
  12. Valentina Ponzo, Ilaria Goitre, Enrica Favaro, Fabio Dario Merlo, Maria Vittoria Mancino, Sergio Riso, and Simona Bo. 2024. Is ChatGPT an Effective Tool for Providing Dietary Advice? Nutrients, 16(4):469. find
  13. Darko Sasanski, Andrej Todorovski, Bojan Trpeski, Dimitar Trajanov, Tome Eftimov, and Riste Stojanov. 2025. Aligning Food Ingredients with Multiple Semantic Resources. In ICT Innovations 2024. TechConvergence: AI, Business, and Startup Synergy, pages 19–33, Cham. Springer Nature Switzerland. find
  14. Qian You, Xuemei Li, Lei Shi, Zhiyong Rao, and Wen Hu. 2025. Still a Long Way to Go, the Potential of ChatGPT in Personalized Dietary Prescrip- find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ranjan-2026-cohere-labs-community-foodbench-qa/