Beans: A Multidisciplinary Benchmark

• Beans are diverse entities that include biological crop organisms (e.g. Phaseolus vulgaris L.), coffee beans, and serve as standardized benchmarks in computer vision, bioacoustics, and astronomy.
• Agricultural studies use beans to evaluate growth responses, stress detection and machine learning performance, with methods reporting up to 93.34% accuracy in classification tasks.
• Technical beans refer to JavaBeans components and distributed data systems, supporting large-scale analyses in astronomy and fostering advancements in web-tier dependency management.

Beans, in the research literature represented here, designate both material objects and technical names. As material objects, the term covers common bean (Phaseolus vulgaris L.), dry bean varieties, Labrador beans, and multiple categories of coffee beans; as technical names, it includes the BEANS bioacoustic benchmark, the BEANS distributed-analysis software package, CAFÉ-BEANS in stellar spectroscopy, and JavaBeans or Managed Beans in JEE web systems. The term therefore spans agriculture, computer vision, robotics, recommender systems, bioacoustics, astronomy, software infrastructure, and Bayesian reasoning (Ghulam et al., 20 Dec 2025, Hagiwara et al., 2022, Hypki, 2016, Shatnawi et al., 2018).

1. Agricultural beans as biological and phenotypic objects

In plant-focused work, beans appear first as crop organisms and phenotype-classification targets. For common bean, Phaseolus vulgaris L., plasma-activated water was tested through seed priming, foliar spraying, and their combination. The laboratory screen showed no germination improvement, but the outdoor pot trial showed notable increases in seedling length, biomass, and antioxidant enzyme activity; SOD, G-POX, CAT, APX, and GR were reported as significantly higher in PAW-treated plants, with seed priming identified as the most effective treatment for plant length and total biomass (Ghulam et al., 20 Dec 2025).

Dry beans also appear as a canonical supervised-learning dataset. The UCI Dry Bean dataset is described as containing 13 properties and seven bean varieties—BOMBAY, SEKER, DERMASON, BARBUNYA, CALI, HOROZ, and SIRA—and the study emphasizes correlated geometric descriptors such as Area, Perimeter, Convex Area, and Major and Minor Axis Length. After standardization and PCA reduction to 8 components, the paper reports that an RBF-kernel SVM achieved Accuracy of 93.34%, Precision of 92.61%, Recall of 92.35%, and F1 Score of 91.40%, outperforming the linear and polynomial SVM variants highlighted in that comparison (Mehta et al., 2023).

Labrador beans appear in a greenhouse pest-detection setting centered on early spider-mite stress. A custom RGBN dataset was acquired with a JAI FS-1600D-10GE camera under real greenhouse conditions, and the main proposed system used a two-stage pipeline: YOLOv8 leaf segmentation followed by a 6-layer sequential CNN for Healthy, Stressed, and Spidermite classification. The best sequential CNN with RGBN reached 90.62% validation accuracy, the average validation-accuracy gain from RGBN over RGB was reported as 6.25% for the usable comparisons, and the best two-stage system improved mAP by 3.6% over the best single-stage end-to-end segmentation baseline (Liu et al., 2024).

These studies collectively position beans as measurable agricultural entities whose growth response, morphology, and stress state can be formalized through biochemical assays, geometric descriptors, and multimodal imaging. A plausible implication is that “beans” function in current research less as a single crop category than as a family of benchmarkable phenotyping and decision-support problems.

2. Coffee beans in vision, quality control, and recommendation

Coffee beans form a distinct technical subdomain because roast level, green-bean quality, origin, and sensory preference are all treated as computationally inferable. In roast-level classification, “Coffee Roast Intelligence” uses photographs of coffee beans captured with an iPhone 12 Mini, collected under LED light-box and natural-light conditions, with some images placing beans in a container or glass bottle. The dataset contains 4,800 total images, with 1,200 images per class for Dark Roasted, Green (Unroasted), Light Roasted, and Medium Roasted. The pipeline applies Gaussian blur, RGB-to-HSV conversion, masking, normalization from $[0,255]$ to $[0,1]$, and augmentation, then fine-tunes MobileNet with batch normalization and dropout; the reported overall accuracy is around 0.82, and the deployment stack includes Flutter on Android, a Flask backend, Firebase, MongoDB, and Azure (Ontoum et al., 2022).

Green coffee beans are also treated as site-dependent color-distribution objects. One study reports that qualified beans from the same growing site have highly similar and unimodal primary-color grayscale distribution curves, whereas defective beans are more variable and often polymodal. The proposed evaluation schemes reduce each bean to either two statistics from one color channel or six statistics across RGB channels and classify with SVM. For the six-feature scheme, accuracies at a 40/60 train-test split were 92.22% for site 1, 84.72% for site 2, 87.78% for site 3, and 98.09% for site 4, and multi-site discrimination among qualified beans from sites 1–3 reached 87.29% (Tan et al., 2024).

Coffee beans further appear as recommendation objects in a service-robot setting. Using 1340 digitized reviews from the Coffee Quality Institute database, with 1312 arabica and 28 robusta samples, one study divides bean information into objective features—such as species, origin, region, variety, color, defects, processing method, and moisture—and subjective sensory qualities including Aroma, Flavour, Body, Sweetness, Acidity, Balance, Uniformity, and Aftertaste. The best supervised predictor of subjective qualities from objective properties was Random Forest with average RMSE 0.3562, and the subsequent k-nearest-neighbour recommendation stage reached up to 92.7% recommendation accuracy when only 10% of beans required predicted sensory features (Berardinis et al., 2020).

Taken together, these results treat coffee beans as computational proxies for roast state, site identity, and sensory expectation. This suggests that coffee-bean research increasingly links vision and metadata not merely to classification, but to operational decisions in roasting, quality assurance, traceability, and human-robot interaction.

3. BEANS as a benchmark for animal-sound machine learning

In bioacoustics, BEANS denotes a public benchmark suite designed to standardize evaluation across animal-sound tasks. The benchmark contains 12 datasets total: 5 classification datasets, 5 detection datasets, and 2 auxiliary datasets from adjacent domains. The classification task is single-label, multi-class and uses accuracy, with the paper defining

$A = \frac{\sum_c ({\it tp}_c + {\it tn}_c)}{CN},$

while detection is framed through sliding windows on long recordings and evaluated with mean average precision. The benchmark standardizes splits, audio preprocessing, and baseline code, and explicitly discourages collapsing all results into one averaged score across datasets and tasks (Hagiwara et al., 2022).

The baseline study reports that VGGish is the strongest overall baseline, pretrained ResNets outperform randomly initialized ResNets on many tasks, SVM is unexpectedly strong on several classification datasets, and detection remains far from solved. The hardest tasks are identified as CBI for classification and RFCx and gibbons for detection. This establishes BEANS as a diagnostic suite rather than a single-number leaderboard (Hagiwara et al., 2022).

Later work uses BEANS as the main downstream testbed for larger pretrained audio models. In one comparison, BioMamba—a Mamba2-based self-supervised model with a wav2vec-2.0-style CNN frontend and Mamba encoder—was fine-tuned on the BEANS tasks and compared against AVES and other baselines. BioMamba was weaker than AVES on 4 of the 5 listed classification datasets, stronger on 4 of the 5 listed detection datasets, and consumed around 40% less VRAM than AVES in the reported inference-memory experiment, supporting the narrower claim of comparable performance with substantially lower memory use rather than uniform superiority (Tang et al., 3 Dec 2025).

A later comparative review restricts BEANS evaluation to the five classification subsets and uses frozen encoders with linear or attentive probing. In that setting, BEATs$_{NLM}$, the extracted audio encoder from NatureLM-audio, achieved 98.57 AUROC and 83.65 Top-1 accuracy on BEANS with attentive probing, ahead of BirdMAE and BEATs. The review argues that, on BEANS-like cross-taxon clip-level classification, broad general-audio self-supervised pretraining transfers extremely well, especially when transformer patch-level features are extracted with attentive probing rather than only pooled embeddings (Schwinger et al., 2 Aug 2025).

4. BEANS as software package and survey acronym

BEANS is also the name of a web-based software package for distributed Big Data storage, querying, analysis, and visualization. Its architecture uses a server/thin-client model with REST communication, Apache Cassandra for distributed storage, Elastic for search, and Pig Latin on Hadoop for analysis. Data are organized into datasets, tables, and notebooks; notebook entries can be text in Markdown, Pig Latin queries, and plots; and plugins extend the system through UDFs and domain-specific integrations such as MOCCA simulation analysis and Gaia Archive access (Hypki, 2016).

The software is motivated primarily by astronomy, especially workflows involving many simulation realizations or large observational archives. Its stated strengths are centralized metadata-rich storage, scalable querying and aggregation, and notebook-driven comparative analysis across many datasets. At the same time, it is explicitly not a machine-learning system and not intended for instant or near-real-time analytics; its preferred mode is supervised, incremental analysis of large tabular datasets (Hypki, 2016).

A separate astronomical use is CAFÉ-BEANS, one of five optical spectroscopic surveys in a study of the interstellar medium toward Galactic O stars. CAFÉ-BEANS is described as the northern, multi-epoch, high-resolution spectroscopic-binarity survey within that ecosystem. The paper is explicit that CAFÉ-BEANS is not a standalone ISM survey; rather, it contributes high-resolution spectra used for DIB measurements, saturation corrections, cloud-component separation, and individual-sightline analysis, complementing GOSSS-based extinction fits (Apellániz, 2015).

Within the combined survey framework, the study reports that the Maíz Apellániz et al. (2014) extinction-law family fits Galactic optical+NIR extinction significantly better than CCM or F99, that most Galactic sightlines have $3.0 \lesssim R_{5495} \lesssim 3.5$, and that $W(5797)/W(5780)$ is anticorrelated with $R_{5495}$. The important point for the term itself is that “BEANS” here is an acronym embedded in a survey network, not a reference to agricultural beans (Apellániz, 2015).

5. JavaBeans and Managed Beans in web-tier dependency structure

In JEE web applications, “beans” refers to reusable software components rather than physical objects. A JavaBean is presented as a regular Java class that is serializable, has a no-argument constructor, and exposes state through public getter and setter properties. Managed Beans are described as “a regular JavaBeans component model that is designed for JSF technology,” with additional framework registration and use in validation, event handling, and data processing (Shatnawi et al., 2018).

The paper’s central concern is dependency implementation in the web tier. In JSP, bean access is made explicit through <jsp:useBean>, <jsp:setProperty>, and <jsp:getProperty>, where the page declares a bean class, assigns an id, and accesses properties through setter or getter conventions. The scope attribute is one of four values—page, request, session, or application—so bean use also carries lifecycle semantics, not just referential semantics (Shatnawi et al., 2018).

Expression language adds a second access mechanism. The paper gives property and method forms such as ${student.firstName}</code>, <code>#{trader.buy}</code>, and <code>#{trader.buy(&#39;SOMESTOCK&#39;)}</code>, showing that pages may depend on beans either by direct JSP declaration or by name-based EL resolution. For JSF, Managed Beans can be registered in <code>faces-config.xml</code> or by <code>@ManagedBean(name = &quot;YouCanUseME&quot;)</code>, and inter-bean injection can be expressed through <code>@ManagedProperty(value=&quot;#{message}&quot;)</code> (<a href="/papers/1803.05253" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Shatnawi et al., 2018</a>).</p> <p>A common misconception is that Managed Beans constitute a wholly separate component model. The paper instead treats them as JavaBeans adapted to JSF: they retain no-argument construction and getter/setter conventions, but are integrated into JSF through configuration and EL-based method or property binding (<a href="/papers/1803.05253" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Shatnawi et al., 2018</a>).</p> <h2 class='paper-heading' id='beans-as-abstract-reasoning-objects-and-granular-models'>6. Beans as abstract reasoning objects and granular models</h2> <p>Beans also serve as abstract tokens in formal inference. In D’Agostini’s analysis of Peirce’s “large bag of beans” example, a bag contains black and white beans in an unknown proportion, with a uniform prior on the white proportion $p \in [0,1]$. After observing$n_W$whites and$n_B$$blacks, the posterior predictive probabilities are</p> <p>$$[0,1]$0

This yields $[0,1]$1, but after 1010 black and 990 white observations the next draw is only approximately a toss-up. The paper’s conclusion is that Peirce’s “balancing reasons” rule fails because the evidential weight of each observation is not constant once the hidden composition parameter is uncertain and updated from the data (D'Agostini, 2010).

In granular mechanics, coffee beans appear again as shape models rather than food commodities. A voxelization-based post-processing method for DEM packings uses a coffee-bean-shaped multisphere particle as Case 3. The bean assembly is poured under gravity into a $[0,1]$2 container, analyzed with normalized voxel size 0.1 based on half the width of the bean, and reported to have average packing fraction $[0,1]$3. The text states that coffee beans show minimal wall effect, nearly flat packing-fraction profiles along $[0,1]$4 and $[0,1]$5, and low packing fraction near walls and corners, with values close to 0.1 at the extremes (Pola et al., 2021).

Coffee beans also appear in tactile robotics as representative enclosed particles. One study on touch sensing for solid particles inside containers names rice and coffee beans as examples and reports estimation of content mass, content volume, particle size, and particle shape over 37 daily particles, with errors of 1.8 g for mass, 6.1 ml for volume, 1.1 mm for size, and 75.6% accuracy for particle-shape estimation (Guo et al., 2023).

Across these uses, beans function as convenient units for expressing uncertainty, packing geometry, and tactile perception. This suggests a broader technical role for the term: beans are repeatedly used where discrete entities must be counted, classified, sensed, or abstracted into features, whether the domain is Bayesian induction, granular media, robotics, or crop science.