COOkeD: Diverse Interpretations in Research

Updated 7 July 2026

COOkeD is a polysemous label encompassing methods for visual state recognition, textual inference, healthcare data transformation, OOD detection, and combinatorial coloring strategies.
It spans diverse applications from AI-driven cooking state analysis for food images to formal methods in combinatorics and ethical evaluations in military-human interactions.
The term underscores transformation and controlled input processing, leveraging advanced models like ViTs, GRUs, and ensemble classifiers to achieve robust and specialized outcomes.

In the cited literature, COOkeD is not a single method or acronym with a stable definition. It is a reused label across several research programs, each centered on a different notion of cookedness, cooking, or constructions figuratively “cooked up” from simpler ingredients. These uses span cooking state recognition in images, state-change prediction from recipe text, post-distribution authorization of healthcare data via Anti-Adversarial processing, heterogeneous ensemble out-of-distribution detection, and formal colorings derived from Szlam’s Lemma (Khan et al., 2022, Wan et al., 2020, Shang et al., 2024, Humblot-Renaux et al., 30 Jul 2025, Myzelev, 2024).

1. Terminological scope and principal meanings

The term appears in the literature as a domain-specific label rather than as a unified framework. In food perception, it refers to recognizing preparation states such as whole, diced, julienne, grated, juiced, creamy paste, and sliced from images (Khan et al., 2022). In recipe understanding, it refers to predicting state changes such as custard cooked or temperature hot from instructional text (Wan et al., 2020). In healthcare ML, “cooked” data are Anti-Adversarial examples that a model must see at both training and deployment to function normally (Shang et al., 2024). In trustworthy vision, COOkeD denotes “CLIP for OOD detection with some extra knowledge,” a three-member ensemble for OOD detection (Humblot-Renaux et al., 30 Jul 2025). In combinatorics, it refers to colorings “cooked up” by Szlam’s Lemma (Myzelev, 2024).

Use of COOkeD	Definition in the cited work	Representative mechanism
Food vision	Cooking state recognition from images	ViT with ImageNet-21k pretraining (Khan et al., 2022)
Recipe understanding	Joint action recognition and state-change prediction	Shared GRU encoder with two MLP decoders (Wan et al., 2020)
Healthcare protection	Anti-Adversarial “cooked” inputs for authorization control	SSIM-constrained perturbation generation (Shang et al., 2024)
OOD detection	Heterogeneous ensemble of supervised, zero-shot, and probe classifiers	Probability averaging plus entropy scoring (Humblot-Renaux et al., 30 Jul 2025)
Mathematics	Colorings “cooked up” from a red-blue partition and template set	Szlam coloring and ordered Szlam coloring (Myzelev, 2024)

A recurrent misconception is to treat COOkeD as a single acronym or benchmark family. The cited record instead shows a polysemous research label whose meaning depends entirely on disciplinary context. This suggests a shared rhetorical motif—transformation under structured rules—rather than a shared technical lineage.

2. Cooking state recognition in vision and language

In computer vision, COOkeD is closely aligned with Cooking State Recognition (CSR), the task of identifying the preparation state of an ingredient rather than its object class. The CSR formulation evaluated in “Rethinking Cooking State Recognition with Vision Transformers” uses the Cooking State Recognition Challenge Dataset introduced by Jelodar et al., with 5,902 images, 7 cooking states, and 18 object types. Images are resized to 224×224, normalized with sample-wise centering to zero mean and unit standard deviation, and split into 4,106 training, 728 validation, and 1,068 test images. The seven-class taxonomy is limited to preparation states and explicitly does not include labels such as raw, baked, boiled, or fried (Khan et al., 2022).

The key technical claim of that work is that ViT self-attention is well suited to CSR because the task depends on subtle texture and geometry cues with high intra-class similarity across ingredients. Two variants are studied: ViT-B/16 with 12 encoder layers, 12 heads, embedding dimension 768, MLP size 3,072, and about 86M parameters; and ViT-L/16 with 24 layers, 16 heads, embedding dimension 1,024, MLP size 4,096, and about 307M parameters. Fine-tuning ImageNet-21k pretrained weights is decisive: ViT-B/16 trained from scratch yields about 55.4% accuracy without augmentation and 55.1% with augmentation, whereas fine-tuning achieves 93.0% without augmentation and 94.3% with augmentation. On the test split, average precision, recall, and F1 are each 0.94; per-class recall reaches 0.97 for diced and is 0.92 for grated, julienne, and whole. The main confusions are juiced versus creamy paste and whole versus sliced (Khan et al., 2022).

A linguistic counterpart appears in “Action Recognition and State Change Prediction in a Recipe Understanding Task Using a Lightweight Neural Network Model.” There, COOkeD concerns procedural text rather than images. The model uses a shared two-layer GRU encoder with 1600 and 800 units over one-hot token inputs, coupled to two independent MLP decoders of 500 hidden units each: one for verb prediction and one for multi-label state-change prediction. The tasks are separated at the output level but coupled through the shared encoder and the summed loss $L_{\text{total}} = L_{\text{action}} + L_{\text{state}}$ . A bounded tangent-based loss is used in practice, and training uses RMSProp with learning rate 1e-4 for 201 epochs. On the 693-sentence Bosselut et al. benchmark, the model reaches 67% average state-change accuracy versus 55% in Bosselut et al. (2018), and about 81% action accuracy, while using 10K training samples rather than 65,815 carefully filtered recipes (Wan et al., 2020).

Related work broadens cooked-state recognition beyond fixed taxonomies. “One-shot recognition of any material anywhere using contrastive learning with physics-based rendering” introduces MatSim, a synthetic-to-real benchmark for material states that explicitly includes smooth transitions such as raw → cooked → burned, mixtures, and transparent containers. A ConvNeXt-based siamese encoder produces 512D L2-normalized embeddings and, on real-image benchmarks, achieves 0.78 top-1 accuracy on Set 1 subclasses with masking, versus 0.67 for the best OpenCLIP H/14 baseline with cropping. Training on mixtures and transparent vessels improves generalization to real food and drink states (Drehwald et al., 2022).

Taken together, these works establish two distinct but complementary notions of COOkeD in food AI: one grounded in fine-grained visual state discrimination, the other in textual state-change inference. Both are driven by subtle transitions rather than coarse object identity.

3. “Cooked” data as Anti-Adversarial authorization in healthcare

In “DataCook: Crafting Anti-Adversarial Examples for Healthcare Data Copyright Protection,” COOkeD denotes healthcare data that have been deliberately transformed into Anti-Adversarial examples (AntiAdv) before distribution. The central idea is not watermarking in the usual pre-distribution sense, but post-distribution control: models trained on cooked data perform normally only when deployment inputs are also cooked through the same process (Shang et al., 2024).

The formal objective is threefold: minimize performance of the cooked model $f_P$ on raw inputs $D_T$ , preserve performance of $f_P$ on cooked inputs $D_P$ so that it matches a raw model $f_r$ on $D_T$ within tolerance $\epsilon$ , and constrain the transformation so that similarity $d(X_T, X_P)$ remains bounded. In practice, the similarity constraint is implemented as SSIM $\ge 0.8$ between each raw image and its cooked counterpart. AntiAdv examples are generated by minimizing a confidence-maximizing loss rather than maximizing a misclassification loss:

$f_P$ 0

Pseudo-labels are obtained from a surrogate raw-data model $f_P$ 1, and perturbations are optimized with Adam at learning rate $f_P$ 2 until the SSIM condition is met. The cooked model itself is then trained with SGD for 200 epochs, batch size 128, momentum 0.9, and learning rate $f_P$ 3 on architectures including ResNet18, ResNet50, VGG16, and ConvNeXt-T (Shang et al., 2024).

The experimental logic is captured by two metrics: Copyright Protection (CP), defined as $f_P$ 4, and Performance Preservation (PP), defined as $f_P$ 5. Lower CP indicates stronger protection; PP near zero indicates that cooked deployment preserves normal performance. On MedMNIST 2D data, AntiAdv (Pseudo) yields CP −18.78, PP −0.67 for ResNet18, CP −9.63, PP −0.91 for ResNet50, CP −16.11, PP −0.56 for VGG16, and CP −13.81, PP −1.59 for ConvNeXt-T. Random noise is substantially weaker, for example CP −2.65, PP −1.61 on ResNet18. On 3D datasets, AntiAdv often outperforms Adv on both CP and PP in larger datasets; one reported example is ResNet18 AntiAdv (Pseudo) CP −18.67, PP −1.75, versus Adv CP −6.03, PP −23.93 (Shang et al., 2024).

A common misunderstanding is to interpret DataCook as generic robustness augmentation. The paper’s deployment model is stricter: the system is designed so that unauthorized raw test inputs cause the cooked model to fail or degrade sharply, and access to the cooking procedure becomes the authorization mechanism. In this usage, COOkeD denotes controlled input-space transformation as governance infrastructure, not culinary state.

4. COOkeD as heterogeneous ensemble OOD detection

A separate and explicitly acronymic usage appears in “COOkeD: Ensemble-based OOD detection in the era of zero-shot CLIP,” where COOkeD means CLIP for OOD detection with some extra knowledge. The method addresses the observation that OOD detection performance is constrained by the capabilities of any single underlying classifier. Instead of choosing between the classical supervised setting and the zero-shot CLIP setting, COOkeD combines both with a third component, a linear probe on CLIP image features (Humblot-Renaux et al., 30 Jul 2025).

The three ensemble members are: a closed-world classifier $f_P$ 6 trained end-to-end on the in-distribution dataset, a zero-shot CLIP classifier using prompts of the form “a photo of a [cls]”, and a linear probe $f_P$ 7 trained on frozen CLIP image embeddings. Each member outputs softmax probabilities over the ID classes, and the ensemble prediction is the simple average

$f_P$ 8

OOD scores are derived post hoc from this averaged distribution, chiefly by entropy

$f_P$ 9

with MSP also reported. The method is intentionally hyperparameter-free beyond CLIP’s learned temperature $D_T$ 0, and no temperature scaling or extra calibration is applied (Humblot-Renaux et al., 30 Jul 2025).

The implementation is lightweight relative to deep ensembles. The closed-world classifier is trained for 100 epochs with SGD, momentum 0.9, weight decay $D_T$ 1, Nesterov enabled, and cosine annealing; the linear probe trains for 20 epochs with the same optimizer and schedule. Images are processed at 224×224. CLIP variants include ViT-B/32, ViT-B/16, ViT-L/14, and ViT-H/14. The approach is modular, post-hoc, and avoids CLIP fine-tuning (Humblot-Renaux et al., 30 Jul 2025).

Reported results show gains in both ID accuracy and OOD robustness. On CIFAR-100, COOkeD attains 86.16% accuracy, versus 83.30% for the standard ResNet18 classifier, 79.98% for the probe, and 66.50% for zero-shot CLIP. On CIFAR-100-N, which introduces about 40% real label noise in training labels, COOkeD reaches 74.60%, versus 62.22% for the standard classifier. On PatternNet, COOkeD reaches 99.86%, while zero-shot CLIP drops to 70.12% under the paper’s “zero-shot shift” setting. On ImageNet-1K, COOkeD reaches 77.26%, and on the full-spectrum covariate-shift benchmark it reaches 64.67%, versus 48.77% for the standard classifier (Humblot-Renaux et al., 30 Jul 2025).

The associated OOD performance is likewise strong. For example, with MSP scoring on CIFAR-100, COOkeD reports 84.32% AUROC on CIFAR-10 and 93.75% on SVHN, compared with 83.16/81.25 for the standard classifier. The paper further argues that entropy of the averaged probabilities is more robust than MSP, especially on far-OOD data, and that averaging per-model MSP values discards informative disagreement (Humblot-Renaux et al., 30 Jul 2025).

An important correction to casual descriptions is that COOkeD is not a single classifier and not merely a CLIP baseline. It is a heterogeneous ensemble whose strength lies in the complementarity of closed-world supervision, zero-shot language alignment, and probe-level feature reuse.

5. Formal and normative uses of “cooked”

In mathematics, the relevant meaning is explicitly figurative. “Characterization of Colorings Obtained by a Method of Szlam” separates the mechanism of Szlam’s Lemma from its chromatic-number consequence by defining Szlam colorings and ordered Szlam colorings of $D_T$ 2. The starting ingredients are a red-blue partition $D_T$ 3 of $D_T$ 4 in which no blue pair is at unit distance, together with a finite set $D_T$ 5 such that no translate of $D_T$ 6 lies entirely in $D_T$ 7. From this, one “cooks up” a proper coloring by choosing, for each $D_T$ 8, an $D_T$ 9 with $f_P$ 0, yielding the bound

$f_P$ 1

The paper then characterizes ordered Szlam colorings via a dominance property: a coloring is ordered Szlam if and only if there is an ordering of the color set with respect to which the coloring is dominant. The paper includes a one-dimensional periodic example with colors green, yellow, and white, and remarks that a planar 7-color Hadwiger–Isbell tiling construction can be converted into a Szlam coloring (Myzelev, 2024).

A very different figurative use appears in “Bad, mad, and cooked: Moral responsibility for civilian harms in human-AI military teams.” Here, “cooked” refers to otherwise good human operators who are placed in untenable decision-making environments through the processes of replacing human decision making with AI determinations in war making. The chapter contrasts bad apples, who intentionally and deliberately kill civilians, with mad apples, who have lost their ability to navigate decisions in war, and cooked apples, whose ethical agency is compromised by automation bias, vigilance decrement, opacity, time pressure, deskilling, distributed responsibility, and organizational constraints. It proposes two mechanisms for analyzing such situations: Decision Responsibility Probes within a Critical Decision Method, and an AI workplace health and safety framework adapted to cognitive and psychological risks in targeting decisions (Devitt, 2022).

These works show that outside ML for cooking or data transformation, COOkeD and “cooked” can function as conceptual operators. In the mathematical case, the term labels a constructive proof pattern; in the ethical case, it labels a sociotechnical condition in which human agency is distorted by system design.

Adjacent literature, though not always using COOkeD as the system name, provides the broader technical substrate for cooked-state modeling. “Real-Time Cooked Food Image Synthesis and Visual Cooking Progress Monitoring on Edge Devices” introduces an oven-based benchmark with 1,708 cooking sessions across 30 recipes, a 70/10/20 train/validation/test split, chef-annotated basic, standard, and extended edible states, and 224×224 RGB processing. Its generator $f_P$ 2 is a compact FiLM-modulated U-Net with 8.68M parameters, and its learned Culinary Image Similarity (CIS) metric is used both as a training loss and as an inference-time progress signal. On the curated oven dataset, COOkeD reports FID 52.18 and LPIPS 0.2145, compared with FID 75.42 and LPIPS 0.2523 for Pix2Pix-Turbo and FID 153.00 and LPIPS 0.4711 for Pix2Pix. The quantized deployment stack occupies 45 MB and runs synthesis at 1.2 s/frame and CIS inference at about 330 ms per live-target pair on a 5 TOPS NPU (Gupta et al., 21 Nov 2025).

Physical studies of cooked matter show that “cookedness” is not always monotone in simple visual or thermal variables. In cooked lutefisk, rheological measurements with a plate-plate rotational rheometer show that the storage modulus $f_P$ 3 depends primarily on fillet size and period of fishing, not systematically on cooking time, cooking temperature, or salt amount within the tested ranges. Large near-head fillets have $f_P$ 4– $f_P$ 5 Pa in the main season, whereas tail fillets have $f_P$ 6– $f_P$ 7 Pa; early-season lutefisk is softer in both classes. The phase angle remains low, with $f_P$ 8 in series 1 and 2 and $f_P$ 9 in series 3 (Feneuil et al., 2021).

Cooked pasta exhibits a different pattern. “Swelling, Softening and Elastocapillary Adhesion of Cooked Pasta” models cooking as water diffusion into a cylindrical or slab geometry, with diffusion coefficients $D_P$ 0 and $D_P$ 1 for 2 mm diameter spaghetti. The modulus evolves through three regimes: a slow decrease in Regime I, a rapid several-orders-of-magnitude drop in Regime II, and a rubbery plateau in Regime III. The raw modulus is $D_P$ 2, while the saturated modulus is about 135 kPa at $D_P$ 3 and 50.8 kPa at $D_P$ 4. The paper further proposes a ruler-based elastocapillary measurement to estimate optimal cooking time (Hwang et al., 2022).

Functional-food work extends the meaning of cookedness to controlled bioactivity after baking. In “How additive manufacturing can boost the bioactivity of baked functional foods,” 3D food printing is used to vary layers, infill, and baking conditions for cookies enriched with encapsulated grape-skin polyphenols. The best reported setting is 4 layers, 30% infill, baked at $D_P$ 5C for 10 min, yielding 754.89 ± 41.89 μmol TR/gdry DPPH activity and 436.74 ± 20.45 mg GAE/gdry total phenolic content. Relative to free-extract cookies at a central condition, this corresponds to +115% antioxidant activity and +173% total phenolic content (Oliveira et al., 2020).

These studies collectively indicate that COOkeD, when interpreted through food science and machine perception, spans at least three technical levels: visual appearance progression, mechanical and physicochemical state, and task-oriented decision thresholds. This suggests that future uses of the term will likely continue to couple perceptual models with either deployment constraints, physical models, or formal decision rules rather than converging on a single canonical definition.