Organizations: State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China. · University of Chinese Academy of Sciences, Beijing, China. · National Institute for Nutrition and Health, Chinese Center for Disease Control and Prevention, Beijing, China.
Precision healthcare, particularly for conditions like hypertension and cardiovascular disease, necessitates monitoring of dietary sodium intake. However, tracking this is hindered by the prevalence of hidden salt in cooking, such as sodium in soy sauce and ketchup. While recipes offer a valuable data source for dietary analysis, sodium-rich seasonings are frequently omitted or described ambiguously in instructions. To solve this issue, we propose SALT, a Sodium Assessing & Level Tracking framework adopting an RAG framework to assess sodium content in recipes. Our framework first introduces a Joint-Factor Embedding Retrieval module to locate similar recipes with specified sodium content for addressing the lack of contextual references. These retrieved samples provide contexts for subsequent inference. Then we design a structured 4-hop Chain-of-Thought inference module to refine the vague estimation from language models through a multi-step sodium estimation. To facilitate our study, we further construct a recipe dataset SALT54k with 54,151 entries labeled with sodium quantities across 11 common seasonings. Results on SALT54k demonstrate that our method achieves state-of-the-art performance in sodium estimation. Additional real-world validations confirm the effectiveness of our method, demonstrating its potential as a practical solution for AI-assisted precision healthcare.
Figures & tables
Figure 1: Diagram of the sodium intake impacting health and a failure attempt to assess sodium content due to an ambiguous recipe.
Figure 2: The system framework for the sodium assessing & level tracking
Figure 3: An illustration of the 4-hop CoT framework. S , CS and F respectively denotes the set of sodium-rich seasonings, cooking styles and flavors, respectively.
Figure 4: Distribution of SALT54k dataset. (a), (b), (c), and (d) present the distribution among different cooking scenarios, primary flavors, recipe regions, and recipe text length. (e) presents a nested pie chart where the inner and the outer layer respectively represent the distribution of coarse- and fine-grained styles; the dark and light shades of the same color represent the coarse-grained style and subdivided fine-grained style (labeled in Appendix B ). In (f), various colors denote 32 distinct ingredients. The size of the blocks in (e) and (f) represents the recipe amount.
Figure 5: Relevance scores radar chart from retrieval evaluation, calculated as the average of three retrieved recipes respective scores, then averaged among the whole testing set. The ( α,β,γ ) setting in this chart is (0.4,0.4,0.2).
Retrieval Method
Relevance Score
Acc
AwT
Random Retrieval
3.02
26.54
29.90
BM25
8.41
37.18
42.34
bge-m3
13.97
50.69
56.27
bge-m3-recipe
19.36
63.62
71.64
Table 1: Downstream sodium estimation performance under different retrieval methods using the Joint RAG + 4-hop CoT setting. Acc and AwT are reported in %. The best results are bolded.
Settings
Models
Acc
AwT
Salt MAE
Soy Sauce MAE
zero-shot
GLM-5
11.20
19.36
2.90
3.80
Deepseek-V3.2
15.30
23.96
2.81
3.54
GPT-5.2
18.18
21.47
2.64
3.77
Gemini 3.1 Pro
19.52
25.24
2.69
3.77
Qwen3-Max
16.80
20.25
3.04
3.73
Naive RAG
GLM-5
27.14 | 30.03
34.39 | 37.82
2.35 | 2.27
3.14 | 3.06
Table 2: Results of different settings on sodium-rich seasonings estimation tasks. The best results are bolded. Acc and AwT are reported in %. The – | – denotes results from 3-shot | 5-shot.
Model
Setting
Acc
AwT
Salt
SS
Llama-3 Chinese
zero-shot
3.82
8.02
3.65
4.81
Naive 3-shot
14.51
18.65
3.10
3.98
Joint 3-shot
20.04
24.74
2.34
3.34
SFT
29.53
35.62
1.87
2.19
Ours
35.47
43.26
1.69
2.11
Qwen3
zero-shot
6.79
12.91
3.47
4.31
Table 3: Results of locally deployed Llama-3-Chinese and Qwen3-8B models under zero-shot, Naive 3-shot RAG, Joint 3-shot RAG, full parameter supervised fine tuned (SFT), and 3-shot + 4-hop CoT (Ours) settings. SS stands for soy sauce. Acc and AwT are reported in %. Salt and soy sauce are reported as MAE. The best results are bolded.
Retrieval Setting
CoT Setting
GLM-5
Deepseek-V3.2
GPT-5.2
Gemini 3.1 Pro
Qwen3-Max
×
×
11.20
15.30
18.18
19.52
16.80
Random
×
16.38
19.41
21.50
22.15
21.83
Random
4-hop
18.29
20.91
24.08
26.54
23.06
Joint
×
34.21
42.55
47.95
54.89
41.17
Joint
Step-by-step
41.76
46.60
51.49
55.91
43.67
Joint
4-hop
56.40
53.96
57.33
63.62
50.75
Table 4: Ablation results on sodium-rich seasonings estimation tasks. The best results are bolded. Results are reported as accuracy in %. The None Retrieval + None CoT, Joint Retrieval + None CoT, Joint Retrieval + 4-hop CoT setting is respectively identical to zero-shot, few-shot, 4-hop CoT + few-shot setting in the Table 2 .
Model
Beijing
Shanghai
Guangzhou
ZS
Ours
ZS
Ours
ZS
Ours
GLM-5
4.66
3.46
5.79
4.06
3.96
3.01
Deepseek-V3.2
4.92
3.58
5.21
3.84
3.86
2.64
GPT-5.2
4.06
2.89
4.69
3.46
3.24
2.09
Gemini 3.1 Pro
4.18
2.55
4.56
3.13
2.93
1.63
Qwen3-Max
3.96
2.30
4.12
2.56
2.73
1.18
Table 5: Results from real-world validation experiments, reported as total sodium MAE (g). ZS indicates zero-shot settings
Model
Acc
AwT
BERT-base-Chinese-Reg
27.35
31.73
bge-m3-recipe-Reg
30.81
38.02
Ours (Gemini 3.1 Pro, 3-shot)
63.62
71.64
Table 6: Comparison with supervised regression baselines on the testing set. Regression models take the concatenated recipe name, ingredients, and cooking steps as input and predict an 11-dimensional seasoning quantity vector. Acc and AwT are reported in %. The best results are bolded.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Raw Expression Type
Example
Normalization Rule
Explicit Solid Quantity
5g Salt; 1 Tsp MSG
Convert to Teaspoon
Explicit Liquid Quantity
20ml Soy Sauce; 1 Tbsp Oyster Sauce
Convert to Yablespoon
Household Spoon Expression
1 Small Spoon; Half Spoon
Map to Tsp/Tbsp According to Convention
Ambiguous Expression
Proper Amount; A Little
Excluded from Labeled Ground Truth
Non-target Seasoning
Sugar; Pepper
Not Included in the 11-dimensional Label
Appendix
Table 7: Unit normalization protocol for SALT54k labels. Solid seasonings are normalized into teaspoons, while liquid seasonings are normalized into tablespoons.
Figure 6: Relevance scores from retrieval evaluation, calculated as the average of three retrieved recipes respective scores, then averaged among the whole testing set. The x-axis denotes ( α,β,γ ) settings as A: (1,0,0), B: (0.5,0.5,0), C: (0.33,0.33,0.33) and D: (0.4,0.4,0.2).
Model
Version
Window
Parameter
Temperature
Top-p
Locally Deployed
Llama3-Chinese
Instruct
8k
8B
1.0
1.0
Qwen3
-
128k
8B
1.0
1.0
API-Invoked
GLM-5
Think
200k
-
1.0
0.95
Deepseek-V3.2
Thinking
128k
-
1.0
1.0
Appendix
Table 8: Overview of Benchmark LLMs. We do not set a maximum token limit across models to ensure complete outputs for the 4-hop CoT process.
Settings
Models
MSG
CE
LSS
DSS
zero-shot
Llama-3-Chinese
0.72
0.47
2.41
1.35
GLM-5
0.68
0.43
2.23
1.10
Deepseek-V3.2
0.70
0.40
2.39
1.06
GPT-5.2
0.68
0.44
2.24
1.14
Gemini 3.1 Pro
0.64
0.33
1.60
0.87
Qwen3-Max
0.60
0.40
1.73
0.65
Appendix
Table 9: Additional results of LLMs on sodium-rich seasonings estimation tasks. The best results are bolded. Abbreviations stand for: M ono S odium G lutamate ( MSG ), C hicken E ssence ( CE ), L ight S oy S auce ( LSS ), and D ark S oy S auce ( DSS ). The – | – denotes results from 3-shot | 5-shot.
Settings
Models
OS
CS
YBS
CBS
TS
zero-shot
Llama-3-Chinese
0.88
0.45
0.37
0.56
0.09
GLM-5
0.86
0.41
0.32
0.55
0.09
Deepseek-V3.2
0.79
0.40
0.36
0.51
0.09
GPT-5.2
0.79
0.31
0.37
0.59
0.10
Gemini 3.1 Pro
0.44
0.21
0.16
0.46
0.06
Qwen3-Max
0.53
0.33
0.15
0.49
0.06
Appendix
Table 10: Additional results of LLMs on sodium-rich seasonings estimation tasks. The best results are bolded. Abbreviations stand for: O yster S auce ( OS ), C hili S auce ( CS ), Y ellow B ean S auce ( YBS ), C hili B ean S auce ( CBS ), and T omato S auce ( TS ). The – | – denotes results from 3-shot | 5-shot.
Figure 7: Examples provided include the input recipe text, the corresponding labels, and the predicted tuples. The red font denotes the incorrect part of the prediction.
Using food package images to monitor sodium and salt content against South Africa's R214 sodium limits is challenging when screening decisions require product identity, nutrition facts panel evidence, reporting basis, and category-specific thresholds. This study presents a conservative image-based workflow that combines region detection, optical character recognition (OCR), product identity and sodium evidence extraction, R214 category assignment, deterministic threshold comparison, and independent vision language model comparison. The evaluation used 442 packaged food products and 3 929 full package images from a real-world South African food packaging dataset. A YOLO26s small detector generated 4 195 region crops, and strict post-processing produced one sodium evidence row per product. The integrated workflow produced 290 OUTSIDE R214 SCOPE, 139 REVIEW, seven SCREEN-PASS, and six SCREEN-FAIL outcomes. The independent Qwen2.5-VL 7B vision language model workflow produced 387 OUTSIDE R214 SCOPE, 31 REVIEW, twenty SCREEN-PASS, and four SCREEN-FAIL outcomes. The workflows agreed on exact R214 category assignment for 415 of 442 products (93.9%) and on whether the assigned category was within R214 scope for 416 of 442 products (94.1%). Final screening outcome agreement was 307 out of 442 products, or 69.5%. Manual verification on 60 products showed lower strict outcome agreement than regulated status agreement, while all manual INSUFFICIENT DATA cases were kept out of SCREEN-PASS and SCREEN-FAIL by both automated workflows. The findings show that conservative image-based screening can organise package evidence, identify clear cases, and assign uncertain cases to REVIEW rather than forcing SCREEN-PASS or SCREEN-FAIL decisions.
Mayimunah Nagayi, Alice Scaria Khan, Tamryn Frank +2
Department of Computer Science University of the Western Cape Cape Town, South Africa · School of Public Health University of the Western Cape Cape Town, South Africa · Department of Dietetics and Nutrition University of the Western Cape Cape Town, South Africa +1
Accurate nutrient estimation from unstructured recipe text is an important yet challenging problem in dietary monitoring, due to ambiguous ingredient terminology and highly variable quantity expressions. We systematically evaluate models spanning a wide range of representational capacity, from lexical matching methods (TF-IDF with Ridge Regression), to deep semantic encoders (DeBERTa-v3), to generative reasoning with large language models (LLMs). Under the strict tolerance criteria defined by EU Regulation 1169/2011, our empirical results reveal a clear trade-off between predictive accuracy and computational efficiency. The TF-IDF baseline achieves moderate nutrient estimation performance with near-instantaneous inference, whereas the DeBERTa-v3 encoder performs poorly under task-specific data scarcity. In contrast, few-shot LLM inference (e.g., Gemini 2.5 Flash) and a hybrid LLM refinement pipeline (TF-IDF combined with Gemini 2.5 Flash) deliver the highest validation accuracy across all nutrient categories. These improvements likely arise from the ability of LLMs to leverage pre-trained world knowledge to resolve ambiguous terminology and normalize non-standard units, which remain difficult for purely lexical approaches. However, these gains come at the cost of substantially higher inference latency, highlighting a practical deployment trade-off between real-time efficiency and nutritional precision in dietary monitoring systems.
Wei-Chun Chen, Yu-Xuan Chen, I-Fang Chung +1
Institute of Biomedical Informatics, National Yang Ming Chiao Tung University, Taipei, Taiwan · Institute of Health Data Science, Chang Gung University, Taoyuan, Taiwan · Department of Artificial Intelligence, Chang Gung University, Taoyuan, Taiwan +1
Multimodal Large Language Models (MLLMs) are increasingly used for dietary assessment from meal images, where retrieval-augmented grounding was shown to sharpen nutrition estimates. However, we find this premise no longer holds for current MLLMs. A modern MLLM's direct estimate now matches or surpasses the full retrieval pipeline. This raises a question: if retrieval no longer improves the overall estimate, can it still deliver the two things clinicians value, accurate portions and a traceable, item-by-item record? We pursue this while preserving what matters for clinical adoption: minimal user burden (a single, unannotated meal image), explainability (an auditable record), and privacy (locally hosted inference). We introduce Open-KNEAD, a knowledge-grounded agentic framework for meal nutrition estimation that is training-free and locally deployable. Each decomposed food item is grounded to a Food and Nutrient Database for Dietary Studies (FNDDS) code via selective, nutrient-aware retrieval, composing an auditable per-item record. Across two open MLLM families and three cuisines, Open-KNEAD improves portion estimates over both prior grounding methods and direct estimation in most backbone-dataset settings. An agent-internal recipe-prior step further recovers the invisible cooking-added energy that biases estimates on non-US cuisine. The advantage is largest on the dietitian-verified ACETADA dataset, where the local open agent surpasses the direct portion estimates of two frontier closed models by roughly 30% and 53%, all while keeping every meal image on local hardware. We release the Open-KNEAD framework and its agent-ready FNDDS knowledge base.