Pith. sign in

REVIEW 4 major objections 8 minor 33 references

Current vision-language models can name dishes but cannot estimate their mass, a new benchmark shows.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 07:49 UTC pith:2V2DFG6T

load-bearing objection A useful three-stage food-VLM benchmark with a real negative result on mass estimation, but the safety-advice labels and missing SFR numbers undercut the strongest claims. the 4 major comments →

arxiv 2607.08423 v2 pith:2V2DFG6T submitted 2026-07-09 cs.AI

OmniFood-Bench: Evaluating VLMs for Nutrient Reasoning and Personalized Health Advice

Qian Jiang , Zhecheng Shi , Jingpu Yang , Zirui Song , Miao Fang This is my paper
classification cs.AI
keywords vision-language modelsfood computingnutritional reasoningbenchmarksafety-critical health advicemass estimationsemantic-physical gap
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

OmniFood-Bench tests six vision-language models on a three-stage chain: naming ingredients and cooking methods, estimating portion and nutrient weights, and giving disease-specific dietary advice. The paper's central finding is that all models are strong at the first stage and weak at the second, with mass-estimation errors frequently above 50% and sometimes over 100%. Because the third stage depends on the second, the errors propagate: models recommend 'moderate intake' for foods a dietitian would flag as 'avoid' for diabetic or kidney-disease patients. The benchmark's point is that deployment in personalized nutrition requires measuring visual-to-physical reasoning, not just classification accuracy.

Core claim

The paper establishes a 'Semantic-Physical Gap': while models achieve near-human accuracy in naming dishes and listing ingredients, their estimates of grams of food and macronutrients have mean absolute percentage errors from roughly 40% to over 400%, and their safety-critical dietary recommendations for patient profiles like diabetes and kidney disease reach at most 46% accuracy. The authors argue the bottleneck is the visual-to-mass estimation step, which propagates error downstream, and that models frequently produce polite advice for dangerous foods, which they call a safety hallucination.

What carries the argument

The central mechanism is the benchmark itself: a curated set of 1,208 food images organized into three progressive tasks — Basic Perception (cooking method classification and ingredient match rate), Quantitative Estimation (portion size and macronutrient mean absolute percentage error), and Safety-Critical Advisory (A/B/C intake-level classification for patient profiles). The paper names the observed dissociation between semantic recognition and physical quantification the 'Semantic-Physical Gap' and attributes it to scale ambiguity and the visual invisibility of intrinsic nutrients, such as sugar in a glaze.

Load-bearing premise

The nutrient weights and intake-level labels used as ground truth come from the source dataset and the authors' chosen A/B/C ranges; if those weights are estimates rather than laboratory measurements, or the ranges are not clinically accepted, every reported error and accuracy inherits the error.

What would settle it

Laboratory-analyze the same dishes and rescore the models: if the mass errors shrink dramatically or the advisory rankings flip, the apparent failure is partly an artifact of the ground truth. Alternatively, run the same benchmark with explicit scale-reference prompts on the same images: if mass error falls below 20%, the visual-to-mass step is not inherently beyond current models.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any vision-language model used as a dietary assistant must be evaluated on physical quantification, not just recognition, before deployment.
  • The three-level chain offers a reusable template for safety evaluation in other high-stakes visual reasoning domains.
  • Because mass estimation is the bottleneck, progress in single-image mass regression or reference-object calibration would directly improve advisory accuracy.
  • Open-weight models are closing the gap on visual grounding, but none shows adequate medical logic for chronic-disease advice.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The intake-level labels were defined by the authors' chosen numerical ranges; a clinical validation against dietitian consensus could shift the measured accuracy rates.
  • A testable extension: adding explicit scale references to prompts might isolate scale ambiguity from other failure modes and show whether mass estimation can be fixed without new architecture.
  • The pattern of polite advice for dangerous foods suggests a conflict between helpfulness-oriented alignment and medical safety; mechanistic study of that tradeoff is a natural next step.
  • If mass estimation is the bottleneck, a system with a depth sensor or known container volume could reach much higher advisory accuracy without stronger reasoning.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. The paper introduces OmniFood-Bench, a three-stage benchmark built from a curated subset of MM-Food-100K to evaluate VLMs on (I) basic perception (cooking method classification and ingredient match), (II) quantitative estimation (portion-size and macronutrient weights), and (III) safety-critical advisory (A/B/C intake decisions for four chronic conditions). Six VLMs — gpt-5.1, gemini-3-flash, claude-sonnet-4, qwen3-vl-8B, InternVL3 5-8B, and Llama-3.2-11B-Vision — are evaluated zero-shot. The authors report strong recognition accuracy but high MAPE in mass/nutrient estimation and low accuracy in disease-specific advice, which they interpret as a 'Semantic-Physical Gap' and a safety hazard for autonomous dietary agents. They identify visual-to-mass estimation and 'medical logic' misalignment as the main bottlenecks and recommend neuro-symbolic integration as future work.

Significance. If validated, the benchmark fills a genuine gap: existing food benchmarks largely stop at classification, whereas OmniFood-Bench explicitly couples perception with physical quantification and downstream health advice. The task hierarchy is sensible, the coverage of restaurant/home/packaged/raw foods is useful, and the qualitative case study of sweet-and-sour pork is instructive. The paper also provides code/data and evaluates both open and closed models in a zero-shot setting. The main empirical trend — high MAPE in Table II and low accuracies in Table III — is visible even without inferential statistics. However, the safety-specific conclusions currently rest on author-defined intake thresholds of unstated clinical provenance and on a Severe Failure Rate that is never reported, so the paper's strongest claims need additional support before OmniFood-Bench can be regarded as a rigorous safety standard.

major comments (4)
  1. [§III-A and Table III] The A/B/C intake-level labels are the ground truth for Task III, and the conclusion that VLMs are unsafe (IV-E) rests on them. Section III-A states labels were assigned 'based on the specific weights (in grams) ... found in the original dataset' and on 'precise numerical ranges' defined by the authors, but no clinical guideline, expert validation, or exact thresholds are given. If MM-Food-100K gram weights are estimates rather than laboratory measurements, or if the thresholds are not clinically valid, every Task III accuracy and the 'Safety Hallucination' conclusion is ungrounded. Please provide the exact thresholds per condition, justify them from a cited clinical source or expert panel, and include a sensitivity analysis over plausible threshold perturbations; also document the provenance/accuracy of the MM-Food-100K nutrient weights.
  2. [§IV-E] The Severe Failure Rate (SFR) is invoked as a highlight of safety risk, but no definition, formula, or numerical values are reported anywhere. As the most safety-relevant quantity, its absence makes the safety claim unauditable. Define SFR precisely, report it by model and patient profile with the same sample sizes as Table III, and state its relationship to the 'sycophantic safety hallucinations' assertion.
  3. [§IV-A and Tables I–III] The comparison mixes evaluation sets: open-source models are scored on 1,208 samples while closed-source models are scored on a 496-sample subset (III-A). No confidence intervals, standard errors, or significance tests are reported, and some headline differences are small (e.g., qwen3-vl-8B 49.56 vs gpt-5.1 37.61 on Homemade Portion in Table I; 44.78 vs 45.56 on Lipids in Table III). These differences may be sampling noise. Report per-cell sample sizes and Wilson confidence intervals (or equivalent), and either add tests or describe all results explicitly as descriptive.
  4. [§IV-E and Fig. 5] The diagnosis that 'medical logic' is the bottleneck is supported only by an anecdote (gemini-3-flash and the burger) and aggregate bar charts. To establish that nutrient knowledge is not the limiting factor, compute advice accuracy conditional on whether the model's nutrient estimate is correct (e.g., within a tolerance of ground truth) and whether it is incorrect. If advice accuracy is similar in both cases, the medical-logic claim holds; if it tracks nutrient accuracy, the bottleneck is upstream. This conditional analysis is needed for the central 'Reasoning Bottleneck' claim.
minor comments (8)
  1. [Abstract and §V] The anonymous code link in the abstract (anonymous.4open.science/r/OmniFood-Bench-7D0B) differs from the GitHub URL in the full text (github.com/PbRQianJiang/OmniFood-Bench). Use one stable link.
  2. [Fig. 4] The y-axis is labeled 'Accuracy (%)' but Table II reports MAPE. Clarify whether accuracy is 1−MAPE or another transform of the error.
  3. [Table I and §III-B] The 'Portion' column is vague; text refers to 'Ingredient Match Rate', 'portion element recognition', and 'ingredient portions'. Define the metric precisely in Section III-B, including how partial matches are scored.
  4. [§III-B] Task III writes P ∈ {Diabetes, Obesity, ...}, but Table III also includes Lipids and Kidney. Specify the full set of patient profiles.
  5. [References] Reference [19] duplicates [11] and contains a typo ('retrival'); reference [25] is a secondary source for GPT-5.1 capabilities — consider citing the primary model documentation if available.
  6. [Table III] Section IV-E calls 46.11% 'barely better than random guessing' in a 3-class problem; the random baseline is 33.3%. Consider reporting per-class accuracy or confusion matrices to contextualize the result.
  7. [Table II] MAPE can exceed 100 and is sensitive to outliers. Consider also reporting median absolute percentage error (MdAPE) or a winsorized MAPE, and state whether the reported values are arithmetic means.
  8. [§III-A] 'Manual spot checks' is the only annotation-quality check described. State how many samples were spot-checked and, if available, inter-annotator agreement.

Circularity Check

0 steps flagged

No significant circularity: the paper is an external zero-shot benchmark evaluation; its conclusions are measured, not derived from fitted inputs or load-bearing self-citations.

full rationale

OmniFood-Bench is a benchmark evaluation rather than a derivation chain. Model outputs for Tasks I-III are produced zero-shot and scored against ground-truth annotations from the external MM-Food-100K dataset; no parameter is fitted to the reported accuracy/MAPE values, so no 'prediction' is forced by construction. The Task III A/B/C labels are author-defined in Sec. III-A ('based on the specific weights (in grams) of proteins, fats, and carbohydrates found in the original dataset' and 'defining precise numerical ranges'), but the evaluated models never see or optimize against those thresholds, so the measured failure rates do not reduce to the labeling rule. The real caveats are evidentiary, not circular: the MM-Food-100K gram weights and the author-chosen intake thresholds lack clinical validation, and Sec. IV-E asserts a 'Severe Failure Rate (SFR) analysis' without reporting SFR numbers. These affect the validity of the safety conclusion, but they are not cases where an output equals an input by definition. The self-citations ([8] includes co-author Zirui Song; [24] includes co-author Jingpu Yang) appear only in related-work context and are not load-bearing; the central claim rests on the paper's own tables and figures. No circular step can be exhibited.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The benchmark's central numbers rest on the accuracy of MM-Food-100K's nutrient labels and on the clinical thresholds the authors define for the A/B/C advice categories. Neither is independently verified inside this paper.

free parameters (1)
  • Intake-level label thresholds (A/B/C ranges) = not disclosed (described only as 'precise numerical ranges')
    These thresholds define the ground-truth labels for Task III. They are chosen by the authors from 'authoritative health and hygiene standards' but no concrete values or derivation are given, so they act as hand-set parameters the evaluation depends on.
axioms (2)
  • domain assumption MM-Food-100K nutrient weights and metadata are accurate
    Section III-A uses the source dataset's weights in grams as ground truth for portion and nutrient tasks; no independent verification is provided.
  • domain assumption The A/B/C intake-level schema is clinically valid for the listed conditions
    The 'Normal Intake / Controlled Intake / Avoid' labels are based on author-defined numerical ranges and are not validated against clinician judgments or published guidelines in this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 9319 in / 10128 out tokens · 97541 ms · 2026-08-02T07:49:44.719517+00:00 · methodology

0 comments
read the original abstract

The rapid integration of Large Vision-Language Models (VLMs) into critical infrastructure promises to revolutionize personalized healthcare and dietary management. However, in the domain of food systems, autonomous agents face a unique and persistent challenge: the "Systemic Information Asymmetry" between visual appearance and intrinsic nutritional composition. Existing benchmarks primarily focus on coarse-grained classification tasks, such as food category recognition, which fail to evaluate the intricate reasoning chain required for real-world dietary management -- specifically, the ability to traverse from identifying hidden ingredients to estimating physical mass, and finally synthesizing safety-critical medical advice. In this paper, we introduce OmniFood-Bench, a comprehensive benchmark constructed from the MM-Food-100K dataset. Unlike previous works, OmniFood-Bench evaluates VLMs across three progressive capabilities: Basic Perception (Ingredients & Cooking Methods), Quantitative Reasoning (Portion Size & Nutritional Profiling), and Safety-Critical Advisory (Disease-Specific Recommendations). We evaluate six state-of-the-art VLMs, including gpt-5.1, gemini-3-flash, and qwen3-vl-8B. Our extensive experiments reveal a startling "Semantic-Physical Gap": while models achieve near-human accuracy in naming dishes, they exhibit catastrophic failure in mass estimation and frequently hallucinate benign advice for high-risk diabetic profiles. This work establishes a rigorous standard for trustworthiness in autonomous agents deployed for public health. The code and datasets are available in: https://anonymous.4open.science/r/OmniFood-Bench-7D0B

Figures

Figures reproduced from arXiv: 2607.08423 by Jingpu Yang, Miao Fang, Qian Jiang, Zhecheng Shi, Zirui Song.

Figure 1
Figure 1. Figure 1: From Visual Recognition to Health Reasoning. Traditional food computing (top) focuses primarily on categorizing dishes. OmniFood-Bench (bottom) introduces a hierarchical evaluation pipeline that requires agents to bridge the “Semantic-Physical Gap”: traversing from ingredient identification to quantitative weight estimation, and finally to personalized, risk-aware medical advisory. tion—specifically target… view at source ↗
Figure 1
Figure 1. Figure 1: From Visual Recognition to Health Reasoning. Traditional food computing (top) focuses primarily on categorizing dishes. OmniFood-Bench (bottom) introduces a hierarchical evaluation pipeline that requires agents to bridge the “Semantic-Physical Gap”: traversing from ingredient identification to quantitative weight estimation, and finally to personalized, risk-aware medical advisory. However, the deployment … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of OmniFood-Bench Data Diversity. The benchmark covers four distinct modalities: Restaurant, Homemade, Packaged Food, and Raw Ingredients, evaluating capabilities from fine-grained weight estimation to precise nutrient extraction. to evaluate ”Biomedical Factual Alignment,” creating a new standard for determining whether an autonomous agent is safe enough to be deployed as a personal health assist… view at source ↗
Figure 3
Figure 3. Figure 3: Holistic Capability Assessment. This radar chart contrasts the performance of six models across 8 dimensions (4 Food Types × 2 Basic Tasks). While models cluster closely on ”Restaurant Cooking” (top), signifi￾cant divergence is observed in ”Packaged Portion” (bottom-left), where open￾source models like Qwen3-VL show competitive performance. C. Results: Quantitative Estimation The transition from semantic r… view at source ↗
Figure 4
Figure 4. Figure 4: Complexity vs. Accuracy. As the number of distinct components in a dish (Variety Level) increases, the portion size estimation accuracy drops sharply for both closed-source (a) and open-source (b) models. This confirms the difficulty of ”Dense Quantitative Reasoning” in food scenes. D. Qualitative Case Study To better understand these failures, we conducted a qualita￾tive analysis of specific error instanc… view at source ↗
Figure 5
Figure 5. Figure 5: Reasoning Bottleneck Analysis. (a) The drop from Portion Recog￾nition to Nutritional Profiling indicates the difficulty of ”Visual-to-Chemical” inference. (b) The inconsistent relationship between Nutrient accuracy and Disease Suggestion accuracy highlights flaws in the ”Medical Logic” capa￾bilities of current VLMs. models excel at recognition but fail at physical quantification, and more critically, they … view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 10 linked inside Pith

  1. [1]

    From autonomous agents to integrated systems, a new paradigm: Orchestrated distributed intelligence,

    K. Tallam, “From autonomous agents to integrated systems, a new paradigm: Orchestrated distributed intelligence,”arXiv preprint arXiv:2503.13754, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    Ai in food sciences and technology– beyond the algorithms,

    L. C. Hoffman and D. Cozzolino, “Ai in food sciences and technology– beyond the algorithms,”Critical Reviews in Food Science and Nutrition, pp. 1–11, 2025

    2025

  3. [3]

    The future of ai in disease detection—a look at emerging trends and future directions in the use of ai for disease detection and diagnosis,

    B. S. Yarman and S. P. S. Rathore, “The future of ai in disease detection—a look at emerging trends and future directions in the use of ai for disease detection and diagnosis,”AI in Disease Detection: Advancements and Applications, pp. 265–288, 2025

    2025

  4. [4]

    Large language model based multi-agents: A survey of progress and challenges,

    T. Guo, X. Chen, Y . Wang, R. Chang, S. Pei, N. V . Chawla, O. Wiest, and X. Zhang, “Large language model based multi-agents: A survey of progress and challenges,”arXiv preprint arXiv:2402.01680, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    A survey of ai agent protocols,

    Y . Yang, H. Chai, Y . Song, S. Qi, M. Wen, N. Li, J. Liao, H. Hu, J. Lin, G. Changet al., “A survey of ai agent protocols,”arXiv preprint arXiv:2504.16736, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    A comprehensive survey of machine learning techniques and models for object detection,

    M. Trigka and E. Dritsas, “A comprehensive survey of machine learning techniques and models for object detection,”Sensors, vol. 25, no. 1, p. 214, 2025

    2025

  7. [7]

    Leveraging artificial intelligence and machine learning to accelerate discovery of disease-modifying therapies in type 1 diabetes,

    M. R. Shapiro, E. M. Tallon, M. E. Brown, A. L. Posgai, M. A. Clements, and T. M. Brusko, “Leveraging artificial intelligence and machine learning to accelerate discovery of disease-modifying therapies in type 1 diabetes,”Diabetologia, vol. 68, no. 3, pp. 477–494, 2025

    2025

  8. [8]

    Benchlmm: Benchmarking cross-style visual capability of large multimodal models,

    R. Cai, Z. Song, D. Guan, Z. Chen, Y . Li, X. Luo, C. Yi, and A. Kot, “Benchlmm: Benchmarking cross-style visual capability of large multimodal models,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 340–358

    2024

  9. [9]

    Towards vlm-based hybrid explainable prompt enhancement for zero- shot industrial anomaly detection,

    W. Cai, W. Huang, Y . Cao, C. Huang, F. Yuan, B. Zhang, and J. Wen, “Towards vlm-based hybrid explainable prompt enhancement for zero- shot industrial anomaly detection,” inProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, 2025, pp. 711– 719

    2025

  10. [10]

    Food-101 – mining dis- criminative components with random forests,

    L. Bossard, M. Guillaumin, and L. Van Gool, “Food-101 – mining dis- criminative components with random forests,” inEuropean Conference on Computer Vision, 2014

    2014

  11. [11]

    Deep-based ingredient recognition for cooking recipe retrieval,

    J. Chen and C.-W. Ngo, “Deep-based ingredient recognition for cooking recipe retrieval,” inProceedings of the 24th ACM international confer- ence on Multimedia, 2016, pp. 32–41

    2016

  12. [12]

    World- cuisines: A massive-scale benchmark for multilingual and multicultural visual question answering on global cuisines,

    G. I. Winata, F. Hudi, P. A. Irawan, D. Anugraha, R. A. Putri, W. Yutong, A. Nohejl, U. A. Prathama, N. Ousidhoum, A. Amrianiet al., “World- cuisines: A massive-scale benchmark for multilingual and multicultural visual question answering on global cuisines,” inProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Asso- ciation f...

    2025

  13. [13]

    Nutrition5k: Towards automatic nutritional understanding of generic food,

    Q. Thames, A. Karpur, W. Norris, F. Xia, L. Panait, T. Weyand, and J. Sim, “Nutrition5k: Towards automatic nutritional understanding of generic food,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 8903–8911

    2021

  14. [14]

    W. Mu, G. A. Kleter, Y . Bouzembrak, E. Dupouy, L. J. Frewer, F. N. Radwan Al Natour, and H. Marvin, “Making food systems more resilient to food safety risks by including artificial intelligence, big data, and internet of things into food safety early warning and emerging risk identification tools,”Comprehensive Reviews in Food Science and Food Safety, vo...

    2024

  15. [15]

    Automatic expansion of a food image dataset leveraging existing categories with domain adaptation,

    Y . Kawano and K. Yanai, “Automatic expansion of a food image dataset leveraging existing categories with domain adaptation,” inProc. of ECCV Workshop on Transferring and Adapting Source Knowledge in Computer Vision (TASK-CV), 2014

    2014

  16. [16]

    Food categorizer: Efficientnet-b1 enhanced cnn for food classification on the food-101 dataset,

    P. C. Shah, J. B. Patel, D. J. Patel, N. V . Patel, N. V . Shah, and S. K. Patel, “Food categorizer: Efficientnet-b1 enhanced cnn for food classification on the food-101 dataset,” inInternational Conference on Information and Communication Technology for Competitive Strategies. Springer, 2024, pp. 355–365

    2024

  17. [17]

    Principles and applications of convolutional neural network for spectral analysis in food quality evaluation: A review,

    N. Luo, D. Xu, B. Xing, X. Yang, and C. Sun, “Principles and applications of convolutional neural network for spectral analysis in food quality evaluation: A review,”Journal of Food Composition and Analysis, vol. 128, p. 105996, 2024

    2024

  18. [18]

    Towards a global spatial-temporal food memory: A vision for privacy-preserving collaborative multimedia analysis,

    Z. Hao, B. Zhang, and H. Li, “Towards a global spatial-temporal food memory: A vision for privacy-preserving collaborative multimedia analysis,” inProceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 12 285–12 294

    2025

  19. [19]

    Deep-based ingredient recognition for cooking recipe retrival,

    C.-w. N. Jing-jing Chen, “Deep-based ingredient recognition for cooking recipe retrival,”ACM Multimedia, 2016

    2016

  20. [20]

    Recipe1m+: a dataset for learning cross-modal embeddings for cooking recipes and food images,

    J. Marin, A. Biswas, F. Ofli, N. Hynes, A. Salvador, Y . Aytar, I. We- ber, and A. Torralba, “Recipe1m+: a dataset for learning cross-modal embeddings for cooking recipes and food images,”arXiv preprint arXiv:1810.06553, 2018

    Pith/arXiv arXiv 2018

  21. [21]

    Capabilities of gemini models in medicine,

    K. Saab, T. Tu, W.-H. Weng, R. Tanno, D. Stutz, E. Wulczyn, F. Zhang, T. Strother, C. Park, E. Vedadiet al., “Capabilities of gemini models in medicine,”arXiv preprint arXiv:2404.18416, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Clicare: Grounding large language models in clinical guidelines for decision sup- port over longitudinal cancer electronic health records,

    D. Li, J. Liang, W. Li, X. Wang, L. Cao, and K. Yu, “Clicare: Grounding large language models in clinical guidelines for decision sup- port over longitudinal cancer electronic health records,”arXiv preprint arXiv:2507.22533, 2025

    arXiv 2025

  23. [23]

    Evaluating the role of faecal calprotectin in older adults: a retrospective observational study,

    R. W. Perry, P. F. Foulser, D. Zhang, P. M. Perez, S. Taylor, A. Sharma, M. Kumaran, S. Balarajah, S. T. Radhakrishnan, R. Sundramoorthiet al., “Evaluating the role of faecal calprotectin in older adults: a retrospective observational study,”The British journal of general practice: the journal of the Royal College of General Practitioners, pp. BJGP–2025, 2025

    2025

  24. [24]

    Personalq: Select, quantize, and serve personalized diffusion models for efficient inference,

    Q. Wang, Q. Guo, Y . Sun, J. Yang, D. Zhang, S. Pang, and Q. Guo, “Personalq: Select, quantize, and serve personalized diffusion models for efficient inference,”arXiv preprint arXiv:2603.22943, 2026

    arXiv 2026

  25. [25]

    Capabilities of gpt-5 across critical domains: Is it the next breakthrough?

    G. P. Georgiou, “Capabilities of gpt-5 across critical domains: Is it the next breakthrough?”arXiv preprint arXiv:2508.19259, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,

    L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qinet al., “A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions,” ACM Transactions on Information Systems, vol. 43, no. 2, pp. 1–55, 2025

    2025

  27. [27]

    Survey and analysis of hallu- cinations in large language models: attribution to prompting strategies or model behavior,

    H. A. Dang, V . Tran, and L.-M. Nguyen, “Survey and analysis of hallu- cinations in large language models: attribution to prompting strategies or model behavior,”Frontiers in Artificial Intelligence, vol. 8, p. 1622292, 2025

    2025

  28. [28]

    How alignment and jailbreak work: Explain llm safety through intermediate hidden states,

    Z. Zhou, H. Yu, X. Zhang, R. Xu, F. Huang, and Y . Li, “How alignment and jailbreak work: Explain llm safety through intermediate hidden states,”arXiv preprint arXiv:2406.05644, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Sg-bench: Evaluating llm safety generalization across diverse tasks and prompt types,

    Y . Mou, S. Zhang, and W. Ye, “Sg-bench: Evaluating llm safety generalization across diverse tasks and prompt types,”Advances in Neural Information Processing Systems, vol. 37, pp. 123 032–123 054, 2024

    2024

  30. [30]

    Mm-food-100k: A 100,000-sample multimodal food intelligence dataset with verifiable provenance,

    Y . Dong, Y . Muraoka, S. Shi, and Y . Zhang, “Mm-food-100k: A 100,000-sample multimodal food intelligence dataset with verifiable provenance,”arXiv preprint arXiv:2508.10429, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Qwen3 technical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Internvl3. 5: Advancing open-source multi- modal models in versatility, reasoning, and efficiency,

    W. Wang, Z. Gao, L. Gu, H. Pu, L. Cui, X. Wei, Z. Liu, L. Jing, S. Ye, J. Shaoet al., “Internvl3. 5: Advancing open-source multi- modal models in versatility, reasoning, and efficiency,”arXiv preprint arXiv:2508.18265, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    The llama 3 herd of models,

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Yang, A. Fanet al., “The llama 3 herd of models,”arXiv e-prints, pp. arXiv–2407, 2024

    2024