Recipe Nutrient Estimation: A Systematic Comparison Between Traditional Methods and LLMs

This paper systematically compares the performance of TF-IDF, DeBERTa-v3, LLMs, and hybrid pipelines on the recipe nutrient estimation task. Key findings: LLMs and hybrid pipelines achieve the highest accuracy under the strict EU 1169/2011 standard, but there is a significant efficiency-accuracy trade-off; traditional methods are fast but have limited accuracy. The study provides guidance for model selection in different scenarios (real-time/precision-first/hybrid) and points out current limitations and future directions.

Accurately estimating nutrient components from unstructured recipe text is a key challenge in dietary monitoring, stemming from two points:

1. Ambiguous ingredient terms: Phrases like "a handful of spinach" or "moderate salt" lack standardization, and processing states (fresh vs. canned tomatoes) also affect nutrients;
2. Variable quantity expressions: Volumes (cups/spoons), weights (grams/ounces), counts (pieces/slices), and ambiguous descriptions (a little) are diverse, requiring complex reasoning for standardization.

Four types of models were evaluated:

1. TF-IDF + Ridge Regression: Lexical baseline, advantages are fast, simple, and interpretable, but cannot handle semantic similarity and complex quantities;
2. DeBERTa-v3: Deep semantic model, expected to have strong semantic capabilities, but performs poorly due to data scarcity and lack of food-specific knowledge;
3. LLM Generative Reasoning: Uses world knowledge, parses ambiguous terms, normalizes units, shows reasoning chains, and has high accuracy;
4. Hybrid LLM Refinement Pipeline: TF-IDF for fast initial estimation + LLM correction, balancing efficiency and accuracy.

The study uses the strict tolerance standard defined by EU Regulation 1169/2011 for evaluation. This regulation specifies accuracy requirements for food nutrition labels, providing a realistic and strict benchmark for the task.

Accuracy Ranking (under EU standard): 1. Hybrid LLM Pipeline > 2. Few-shot LLM > 3. TF-IDF > 4. DeBERTa-v3; LLMs perform best in all nutrient categories. Efficiency Comparison:

Method	Inference Latency	Accuracy
TF-IDF	Millisecond level	Medium
DeBERTa-v3	Hundred-millisecond level	Low
LLM	Second level	Highest
Hybrid Pipeline	Second level	Highest
Conclusion: Higher accuracy comes at the cost of real-time performance.

1. Real-time applications: For example, mobile real-time scanning, choose TF-IDF (millisecond-level response);
2. Precision-first: For example, medical nutrition monitoring, choose LLM (meets regulatory accuracy);
3. Hybrid deployment: TF-IDF for initial estimation in the fast phase, LLM optimization in the background for a progressive experience.

Current Limitations: Evaluated only on FoodBench-QA, no analysis of LLM costs, no multilingual support; Future Directions: Domain-adaptive fine-tuning of LLMs, evaluation of efficient small LLMs, integration with structured nutrition databases, design of interactive systems.