Energy estimates of open-weight large language models across stated food weights: a dose-response evaluation
Authors: Donato Riccio, Fabrizio Maturo, Iulia Brezeanu
Organizations: Department of Engineering and Science, Universitas Mercatorum, Rome, Italy · Department of Economics, Statistics and Business, Universitas Mercatorum, Rome, Italy · Geometry Lab L.L.C-FZ, Dubai, United Arab Emirates
We tested whether energy estimates of open-weight large language models (LLMs) scale proportionally with an exactly stated food weight, and whether computing energy density times weight removes the error. Under a prespecified protocol, seven open-weight LLMs estimated the energy of 168 randomly drawn foods from the USDA Food and Nutrient Database for Dietary Studies at nine stated weights, from 0.25 to 4 times the default amount. The primary outcome was the elasticity of the estimate with respect to weight (1: proportional scaling). A two-step estimate multiplied each model's own stated energy density by the weight in code. All seven models compressed their estimates, with elasticities from 0.548 to 0.979 (every 95% CI below 1); four compressed severely (0.548 to 0.709). These four overestimated a quarter of the default amount by 47% to 130% and underestimated four times it by 20% to 34%. Supplying the density restored near-proportional scaling in three of the four, and the two-step estimate cut the typical error factor at these extremes from 1.47-2.43 to 1.20-1.38. Compression persisted at full precision and grew in five of seven models under a plain question. The models hold the energy densities needed for proportional scaling and answer near a typical serving, reproducing the flat-slope syndrome of dietary recall. Accuracy at usual portions does not reveal it. Food-logging tools should compute the energy of weighed entries as density times weight, and evaluations should test fixed foods across stated weights.
Figures & tables
Model
Developer
Parameters
Architecture
Sampling ( T , p , k )
Qwen3.5-35B-A3B
Alibaba
35B (3B)
mixture of experts
1.0, 0.95, 20
gpt-oss-20b
OpenAI
21B (3.6B)
mixture of experts, reasoning
1.0, 1.0, 0
Gemma-4-26B-A4B
Google
26B (3.8B)
mixture of experts
1.0, 0.95, 64
Mistral-Small-3.2-24B
Mistral AI
24B
dense
0.15, 1.0, 0
Llama-3.3-70B
Meta
70B
dense
0.6, 0.9, 0
Phi-4
Microsoft
14B
dense
1.0, 1.0, 0
Table 1: Models. Parameters are total (active per token for mixture-of-experts models). Sampling gives temperature T , nucleus probability p and top- k ( k = 0: off). Revisions are in Supplementary File 1.
Ratio at
Density
Model
β
95% CI
0.25q
q
4q
Greedy
Ratio
MdAPE
Valid
Qwen3.5-35B-A3B
0.979
0.967–0.990
0.95
0.96
0.91
0.987
0.91
15.0%
100.0%
gpt-oss-20b
0.978
0.955–0.999
1.01
0.96
0.95
0.977
0.98
14.9%
100.0%
Gemma-4-26B-A4B
0.954
0.935–0.968
1.14
1.05
1.00
0.953
1.00
9.2%
99.9%
Mistral-Small-3.2-24B
0.709
0.678–0.741
1.47
1.02
0.67
0.722
1.00
12.1%
99.7%
Llama-3.3-70B
0.694
0.667–0.724
1.68
1.04
0.77
0.689
0.99
10.7%
99.5%
Table 2: Scaling and density knowledge by model on 168 foods. β : elasticity of the energy estimate with respect to the stated weight (1: proportional), with 95% CI from the food bootstrap. Ratio: geometric mean over foods of estimate / reference energy at a stated weight of 0.25q , q and 4q . Greedy: elasticity of the greedy responses. Density: the model’s answer to the density question against the FNDDS value, as the median ratio and the median absolute percentage error (MdAPE) over foods. Valid: share of the 9,408 designed responses that parsed. Density statistics of Gemma-4-26B-A4B use 166 foods, which gave a valid density answer.
Figure 1: Dose-response curves of energy estimates against stated weight on 168 foods. Each panel shows the geometric mean over foods of estimate / reference energy (points) at nine stated weights from a quarter to four times the FNDDS default amount, with a 95% simultaneous food-bootstrap band (shaded). Both axes are logarithmic. Proportional scaling is the dashed horizontal line at 1; a curve that falls from left to right is compressed. Panels are ordered by elasticity. Curves use the foods with valid responses at all nine weights (167 for Mistral-Small-3.2-24B and 162 for Llama-3.3-70B; 168 otherwise).
Elasticity ( 0.25q , q , 4q )
MALE at 0.25q
MALE at 4q
Model
Without
With
Difference
Direct
Two-step
Direct
Two-step
Mistral-Small-3.2-24B
0.718
0.973
0.255 (0.218–0.292)
0.467
0.204
0.518
0.204
Llama-3.3-70B
0.720
0.995
0.274 (0.244–0.305)
0.570
0.185
0.433
0.183
Phi-4
0.703
0.984
0.281 (0.251–0.312)
0.663
0.246
0.386
0.246
Llama-3.1-8B
0.548
0.776
0.227 (0.144–0.303)
0.887
0.319
0.580
0.319
Table 3: Supplied density and two-step estimate in the four severely compressing models, 168 foods. Elasticity over the stated weights 0.25q , q and 4q without and with the FNDDS energy density in the prompt, and the paired difference. MALE: mean absolute log error of the direct estimate and of the two-step estimate (the model’s own density answer times the stated weight) at 0.25q and 4q , over the foods with a valid direct estimate at that weight (all 168, except 165 at 0.25q and 167 at 4q for Llama-3.3-70B). Intervals are 95% CIs from the paired food bootstrap. For Llama-3.1-8B, the three-weight elasticity without the density coincides with its nine-weight elasticity (0.548).
Figure 2: Supplied density and two-step estimate in the four severely compressing models, 168 foods. (A) Elasticity of the energy estimate over the stated weights 0.25q , q and 4q with the weight only and with the FNDDS energy density added to the prompt; the dashed line at 1 is proportional scaling. (B) Mean absolute log error of the direct estimate at 0.25q and at 4q , and of the two-step estimate, whose error over the same foods is the same at every stated weight. Table 3 gives the elasticity differences with 95% CIs; the text gives the paired differences in mean absolute log error.
Figure 3: Elasticity by food group and model, 24 foods per group and nine stated weights per food. Each cell is the elasticity β estimated within the group; 1 is proportional scaling; cells below 1 are shaded red and cells above 1 blue. Models are ordered by overall elasticity. The comparison of groups is exploratory.
Elasticity
Simulated
Model
Median
Sampled
Difference (95% CI)
switching
Between
Weight
Qwen3.5-35B-A3B
0.971
0.979
0.42 (0.33 to 0.52)
−0.30
57%
0.98
gpt-oss-20b
0.991
0.978
1.00 (0.75 to 1.25)
−0.30
51%
1.00
Gemma-4-26B-A4B
0.951
0.954
0.78 (0.58 to 0.98)
−0.28
59%
0.99
Mistral-Small-3.2-24B
0.753
0.709
0.95 (0.80 to 1.10)
−0.85
77%
0.80
Llama-3.3-70B
0.707
0.694
2.37 (2.10 to 2.64)
−1.00
74%
0.76
Table 4: Token probabilities of the energy answers, 168 foods. Median: elasticity of the median of each answer’s probability distribution; Sampled: elasticity of the sampled answers (Table 2 ). Difference: expected log-likelihood of the compromise account minus that of the switching account at 0.25q , 0.35q , 2.83q and 4q , on the prompts whose computed value C and remembered value R differ by at least 50% (597 to 630 per model; 166 foods for Gemma-4-26B-A4B and 167 for Llama-3.1-8B, whose other foods had no remembered value), mean per food with 95% CI (food bootstrap); positive values favor the compromise. Simulated switching: the same difference when the probability of each prompt is replaced by a mixture of the model’s own computed and remembered distributions at the compromise weight fitted to that prompt (Supplementary File 1, Table S14). Between: share of prompts whose distribution median lies between R and C . Weight: median fitted weight of the compromise account on the computed value.
Figure 4: Where the probability of an energy answer lies. (A) Probability distribution of the energy answer of Llama-3.3-70B for one food, cheese spread (American or Cheddar cheese base), at a quarter (blue) and four times (orange, hatched) the default amount, with the remembered value R (median energy of the model’s answer without an amount), the computed values C (the median of the model’s density answer times the weight; dashed) and the reference energies (dotted). The food is the one with the median likelihood difference for this model. (B) Mean probability density by position t between the remembered value ( t=0 ) and the computed value ( t=1 ), over the prompts at 0.25q , 0.35q , 2.83q and 4q whose remembered and computed values differ by at least 50%; line styles and markers identify the models.
Photo-based estimation of food energy could ease dietary assessment, and general-purpose multimodal large language models (LLMs) make such estimates easy to obtain. Most evaluations use one photo taken before eating and compare with energy served, while dietary assessment needs energy eaten. We evaluated eight models from three vendors (Claude Haiku, Sonnet and Opus 5.5; GPT-6 luna, 6.1 sol and 6 astra; Gemini 3.8 Flash and 4 Argon) on 802 meals from ACETADA, a controlled feeding trial with weighed intake, with the before photo only, before and after photos in separate calls (subtraction), or both photos in one request. Three results ran against intuition. Subtraction under-estimated intake for six of eight models, because the models still report hundreds of kilocalories for a nearly empty tray. Scaling the before-only estimate by the average share eaten (0.87) matched much of the gain from the after photo; beyond that constant, the second photo lowered meal-level error by only 2.2 to 4.4 points, mainly in meals with food left. With one photo, two models had a near-zero intake bias only because they under-estimated the served energy and ignored uneaten food. In ablations, adding the after photo improved accuracy under both prompt formats for seven models, while prompt effects varied across models. Over 148 feeding days, five models kept the 90% confidence interval of their mean error within a pre-set 10% margin; one model also met the pre-set 30% limits of agreement, narrowly (-28% to +27%; limits outside 30% cannot be excluded), and the other seven did not. Protein was under-estimated by every model, most in protein-dense meals; repeated calls changed error by at most 1.4 points; a high reasoning effort helped only the smallest model. A low average bias is not evidence of accurate individual intake estimation; such estimates should support, not replace, assessment by a dietitian.
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.
Image-based dietary assessment promises to replace costly, bias-prone manual recalls, but portion estimation remains a major blocker. Multimodal LLMs (MLLMs) recognize a wide range of foods zero-shot in uncontrolled photos, yet they are weak at portion estimation -- a gap we measure across the current frontier (Gemini, GPT, and Claude flagships alike). We present a method that enhances a frozen, commercial MLLM with an accurate portion head: a small geometry-enhanced network on a frozen DINOv2 backbone with a structured softmax-ownership volume, consuming the MLLM's per-food name, bounding box, and density range -- no depth sensor, no MLLM fine-tuning. Evaluated fully open-vocabulary on three real-world benchmarks, the head cuts per-food portion error by 33-41% relative to the MLLM alone, outperforms every flagship MLLM's direct estimates, and surpasses each benchmark's originally published image-only model at its own reported metric.