MANGO: Multifaceted Research Applications

• MANGO is a multifaceted term referring to both the climacteric fruit species Mangifera indica L. and a range of technical artifacts in contemporary research.
• In agricultural contexts, studies on mango ripening reveal significant differences in nutritional composition and cytogenotoxic effects based on different artificial ripening agents.
• In computational research, MANGO frameworks offer modular designs for hyperparameter tuning, agent simulation, navigation benchmarks, multimodal fusion, and structured optimization.

MANGO denotes both mango as the climacteric fruit species Mangifera indica L. and a recurrent acronym family reused across contemporary research for software libraries, benchmarks, optimization methods, multimodal models, and agent systems. In recent arXiv literature, the name has been attached to a Python library for parallel hyperparameter tuning, a modular agent-simulation framework, a benchmark for text-based mapping and navigation, several multimodal fusion and robustness methods, web-navigation and robotic-testing systems, and multiple optimization frameworks for continual learning, hierarchical reinforcement learning, and offline design (Obadiah et al., 2024, Sandha et al., 2020, Schrage et al., 2023, Ding et al., 2024, Truong et al., 13 Aug 2025, Wang et al., 13 May 2026, Awasthi et al., 18 May 2026, Tong et al., 20 Apr 2026, Valle et al., 23 Jun 2026, Zhou et al., 6 Jun 2025).

1. Nomenclature and recurrent acronym use

A common source of confusion is that MANGO is not a single standardized acronym. Different papers expand it differently and use it for unrelated technical artifacts. This suggests that disambiguation depends on the full title, domain, and arXiv identifier rather than on the acronym alone.

Use	Expansion or referent	Domain
MANGO	parallel hyperparameter tuning library	Bayesian optimization software
mango	modular python agent framework	agent-based simulation
MANGO	mapping and navigation benchmark	LLM evaluation
MANGO	Multimodal Acuity traNsformer for intelliGent ICU Outcomes	clinical multimodal prediction
MANGO	Multimodal Attention-based Normalizing Flow Approach to Fusion Learning	multimodal representation learning
MANGO	Multi-Agent Network Gradient Optimization	multi-agent LLM systems
Mango	multi-agent web navigation via global-view optimization	web agents
MANGO	automated multi-agent test oracle generation	VLA testing
ManGO	learning the design-score manifold	offline optimization

This reuse is not merely stylistic. In some cases MANGO denotes an executable library or framework, in others a benchmark, and in others a model family or training method. A closely related orthographic variant, ManGO, is used for a diffusion-based offline optimization framework rather than an all-caps acronym (Sandha et al., 2020, Schrage et al., 2023, Ding et al., 2024, Zhang et al., 2024, Truong et al., 13 Aug 2025, Wang et al., 13 May 2026, Tong et al., 20 Apr 2026, Valle et al., 23 Jun 2026, Zhou et al., 6 Jun 2025).

2. Mango as Mangifera indica L. and as a postharvest toxicology subject

In agricultural and food-science usage, mango refers to Mangifera indica L., a climacteric fruit whose ripening changes pigment, sugar, acid, flavor, aroma, texture, and color. One study centered on mangoes artificially ripened with wood ash, herbaceous ash, and rice chips in Birnin Kebbi, Nigeria, comparing them with naturally ripened fruit and assessing both proximate composition and cytogenotoxicity by the Allium cepa assay (Obadiah et al., 2024).

The experimental design used two naturally ripened mangoes and six fully mature unripe mango fruits, with the unripe fruits divided into three artificial-ripening groups of two mango replicates each. Ripening was judged by a pale yellow appearance and yielding to thumb pressure. For cytogenotoxicity, the study used 20 Allium cepa bulbs, with 5 bulbs each assigned to naturally ripened mango, wood ash, herbaceous ash, and rice chips. Root tips were grown over tap water for five days, then exposed to ripening agents for another five days, with concentrations of 5 g/dl, 10 g/dl, 15 g/dl, 20 g/dl, 25 g/dl, plus control for each artificial agent (Obadiah et al., 2024).

The proximate-composition results reported significant differences at P < 0.05. Wood ash produced the highest moisture content, 81 ± 1, and the lowest carbohydrate content, 9.92 ± 1.1. Rice chips produced the highest ash content, 1.5 ± 0.5, and protein, 0.5 ± 0.01. Natural ripening yielded the highest fibre, 11.46 ± 1.04, and the highest fat, 0.0095. The study also stated that there was no significant difference in carbohydrate among naturally ripened mangoes (14.99 ± 0.26), rice-chip-ripened mangoes (15.4 ± 0.89), and herbaceous-ash-ripened mangoes (14.95 ± 0.61), while wood ash was distinctly lower (Obadiah et al., 2024).

The Allium cepa assay indicated reduced root growth relative to control and multiple chromosomal aberrations. On day 1, wood ash showed the highest root growth (2.62 ± 0.14 cm) and herbaceous ash the least (2.18 ± 0.12 cm). Across concentrations, the control increased from 1.98 ± 0.1 cm initially to 5.45 ± 0.96 cm by day 4, whereas treated groups remained much lower. Reported abnormalities included sticky chromosomes, vagrant chromosomes, laggard chromosomes, and chromosome bridges at telophase. The study identified herbaceous ash as the most toxic ripening agent and rice chips as the least toxic, concluding that induced ripening “could induce toxicities” and recommending public awareness regarding these hazards (Obadiah et al., 2024).

3. Software libraries and simulation frameworks

One major technical use of MANGO is Mango: A Python Library for Parallel Hyperparameter Tuning, a black-box optimizer built around parallel Bayesian optimization with an explicit separation between optimizer and scheduler. Its core workflow has four abstractions: the user defines a hyperparameter search space, provides an objective function, selects a parallel optimization algorithm, and lets Mango use a scheduler to evaluate candidate configurations. The library is intentionally framework-agnostic: it can run in a single-machine serial setting, with local parallelism, or on a distributed cluster such as Celery. Search spaces are Python-dictionary based, support continuous, discrete, and categorical variables, and are compatible with scikit-learn; the library supports all 70+ distributions from scipy.stats, permits user-defined distributions, and includes a built-in loguniform. Its optimization engine uses UCB and provides two parallel algorithms—one motivated by Desautels et al. and one clustering-based method motivated by Groves et al.—plus a random optimizer. The paper emphasizes support for partial results, missing evaluations, out-of-order completion, straggler workers, and faulty workers, and reports performance comparable to Hyperopt while being used in production at Arm Research with Celery on a Kubernetes cluster (Sandha et al., 2020).

A distinct framework, mango: A Modular Python-Based Agent Simulation Framework, targets communication-heavy multi-agent simulations rather than hyperparameter tuning. It supports both inheritance-based agents via mango.Agent and composition-based design via a role system and mango.RoleAgent. Its architecture separates agents, containers, codecs, schedulers, roles, and external connection containers. The paper describes three container implementations—TCP, MQTT, and EC for external connection—together with two codecs, JSON and protobuf. Every agent has a scheduler capable of starting and stopping tasks and determining whether a task has finished or is sleeping; the framework also provides a (distributed) clock mechanism for time synchronization, multiprocessing support through ProcessTasks and separate agent processes, and co-simulation interfaces for systems such as OMNeT++ and mosaik. The design goal is that the same agent implementation can run in a single process, across multiple processes on one machine, or across different machines without changing agent logic (Schrage et al., 2023).

Taken together, these two systems show that MANGO often denotes not just an algorithm but an architectural stance. In one case the key principle is separation of optimizer and scheduler; in the other it is separation of behavior, communication, serialization, and timing. The commonality is modularity rather than a shared codebase or shared domain (Sandha et al., 2020, Schrage et al., 2023).

4. Benchmarks, navigation systems, and testing infrastructure for agents

MANGO has also been used for evaluation and control of agents. MANGO: A Benchmark for Evaluating Mapping and Navigation Abilities of LLMs constructs a text-only navigation benchmark from 53 mazes extracted from Jericho textgames. It evaluates destination-finding (DF) and route-finding (RF), uses both explicit and imputed edges, and contains 21,046 DF skeletons and 14,698 RF skeletons. The paper reports that GPT-4 was the strongest evaluated model but still answered only about half of RF questions correctly, whereas humans answered a sampled set perfectly. It also introduces ANSWERABLE and EASY labels to separate questions that merely require recalling traversed information from those that require reasoning over unseen but valid routes (Ding et al., 2024).

A later system, Mango: Multi-Agent Web Navigation via Global-View Optimization, addresses a related but operational setting: web agents navigating complex websites under a fixed action budget. Rather than starting from the root URL and exploring downward, Mango first builds a coarse global view by lightweight crawling and site-restricted search, scores candidate URLs with BM25, and then formulates URL selection as a multi-armed bandit problem solved with Thompson Sampling. A Beta prior is initialized from the normalized BM25 relevance score, and an episodic memory stores prior trajectories and reflections. On WebVoyager, the paper reports a 63.57% success rate with GPT-5-mini, improving on the best baseline by 7.3 absolute points; on WebWalkerQA, it reports 52.50%, surpassing the best baseline by 26.8 absolute points (Tong et al., 20 Apr 2026).

A third system, MANGO: Automated Multi-Agent Test Oracle Generation for Vision-Language-Action Models, repurposes the acronym for robotic testing rather than navigation. It decomposes a natural-language instruction into atomic tasks, generates a reusable Atomic Task Library, synthesizes simulator-grounded atomic-task oracles, and composes these into executable fine-grained oracles through collaborative Generator, Assessor, and Judge agents. The evaluation on LIBERO_10 and RoboCasa Humanoid Tabletop reports, for oracle generation quality, validity 0.950, executability 0.790, and similarity 0.913 on LIBERO_10, and validity 1.000, executability 0.933, and similarity 0.947 on RoboCasa. For failure detection after simulator-level fixes, the generated fine-grained oracles achieved accuracy 0.91, precision 0.91, recall 0.95, F1 0.93, and localization accuracy 0.78 (Valle et al., 23 Jun 2026).

These uses are unified by their concern with structured reasoning over environments: map induction in text worlds, budget allocation over candidate web entry points, and decomposition of robotic tasks into executable intermediate checks. What differs is the formal object being optimized—paths, URLs, or oracle structure—not the high-level emphasis on stepwise rather than purely end-state reasoning (Ding et al., 2024, Tong et al., 20 Apr 2026, Valle et al., 23 Jun 2026).

5. Multimodal, vision, and media architectures

Several papers use MANGO for multimodal modeling in the strict representation-learning sense. MANGO: Multimodal Attention-based Normalizing Flow Approach to Fusion Learning proposes an explicit, interpretable, and tractable multimodal fusion model based on normalizing flows and a new Invertible Cross-Attention (ICA) layer. It defines three partitioning mechanisms—MMCA, IMCA, and LICA—and uses a latent compression stage before the flow model. The default implementation stacks $L=12$ cross-attention blocks. On semantic segmentation, it reports Pixel Acc. 81.5, mAcc. 71.6, and mIoU 59.2 on NYUDv2, and 83.9 / 67.2 / 54.1 on SUN RGB-D; on MM-IMDB, it reports Micro-F1 71.7 and Macro-F1 68.2 (Truong et al., 13 Aug 2025).

In clinical prediction, MANGO: Multimodal Acuity traNsformer for intelliGent ICU Outcomes fuses EHR data, wearable accelerometer data, video-derived facial action units, and ambient sensor data from the ICU-Multimodal dataset collected from 310 adult patients in nine specialized ICUs. Its fusion mechanism is masked multi-head self-attention, allowing modalities to be absent rather than imputed. The best-performing model reports AUROC 0.76 (95% CI: 0.72–0.79) for predicting transitions in acuity status and need for life-sustaining therapy, and AUROC 0.82 for acuity-status prediction (Zhang et al., 2024).

In robustness research, Mango was introduced as a Multimodal Adversarial Noise GeneratOr for vision-and-language pre-trained models such as UNITER and LXMERT. Instead of iterative PGD, it learns lightweight generators that produce adversarial perturbations directly in the embedding spaces of image regions and word tokens. The paper evaluates four robustness dimensions—linguistic variation, logical reasoning, visual content manipulation, and answer distribution shift—and states that Mango achieves new state of the art on 7 out of 9 robustness benchmarks, while also being about 25% faster than Villa in comparable settings (Li et al., 2020).

Vision and graphics papers reuse the name differently again. Deep Mangoes studies RGB-image detection and cultivar identification of mango fruits in orchard imagery using Faster R-CNN on tiled 500 × 500 crops derived from 4000 × 6000 images, obtaining F1-score = 0.90 for tile-level fruit detection and F1-score = 0.56 for simultaneous detection and cultivar identification, with detected-fruit cultivar identification rates in the 80–90% range (Borianne et al., 2019). MANGO: A Mask Attention Guided One-Stage Scene Text Spotter eliminates RoI operations by using Position-aware Mask Attention with instance-level and character-level masks; on Total-Text, it reports 72.9 under None and 83.6 under Full, and on CTW1500 it reports 58.9 under None and 78.7 under Full (2012.04350).

The acronym also appears in generative and sim2real media systems. MANGO:Natural Multi-speaker 3D Talking Head Generation via 2D-Lifted Enhancement combines a diffusion-based transformer, dual-audio interaction, a 3D Gaussian Renderer, and the MANGO-Dialog dataset with over 50 hours of aligned 2D-3D conversational data across 500+ identities (Zhu et al., 5 Jan 2026). MANGO as Multiview Augmentation with Novel Generated Observations is an unpaired sim2real image-translation method for fixed-camera robot datasets; it adds a segmentation-conditioned InfoNCE loss, a modified PatchNCE loss, and a highly regularized discriminator, and the paper states that policies trained on MANGO-augmented data can reach success rates as high as 60% on shifted views where the non-augmented policy fails completely (Coholich et al., 14 Jan 2026).

6. Optimization and learning frameworks

A final cluster of uses centers on adaptive optimization. MANGO: Meta-Adaptive Network Gradient Optimization for Online Continual Learning addresses the stability–plasticity tradeoff in single-pass online continual learning. Its two core mechanisms are gradient-gating, which scales parameter updates according to sensitivity, and meta-learned regularization, which adapts layer-wise stability coefficients using replay as a forgetting evaluator. On Split CIFAR-100 with a 2000-sample buffer, the paper reports Acc 19.72 ± 1.01%, AAA 31.53 ± 0.37%, and WC-Acc 12.69 ± 0.55%; on CLEAR-10, it reports Acc 66.91 ± 1.20% and positive BWT +15.12 ± 3.3% at the same buffer size (Awasthi et al., 18 May 2026).

In hierarchical reinforcement learning, MANGO: Multi-layer Abstraction for Nested Generation of Options decomposes long-horizon sparse-reward problems into multiple abstraction layers via concept functions, abstract actions, and nested options. Higher-layer options invoke lower-layer options rather than primitive actions directly, and training proceeds layerwise from the bottom up. Experiments on procedurally generated $16 \times 16$ grid environments show that well-tuned MANGO can achieve near-perfect completion earlier than Q-learning in some settings, while also exposing a major limitation: poor lower-layer learning propagates upward and can destabilize the hierarchy (Arcudi et al., 25 Aug 2025).

A related but distinct variant, ManGO, learns a design-score manifold for offline optimization with a diffusion model over joint design–score variables rather than using a separate surrogate or inverse model. It supports both single-objective and multi-objective optimization, uses score-aware sample reweighting, and adds adaptive inference-time scaling through self-supervised importance sampling. On Design-Bench single-objective tasks, the paper reports mean ranks of 2.2 / 24 for ManGO and 1.4 / 24 for ManGO + self-IS; on multi-objective benchmarks it reports top mean ranks such as 2.0 (HV) and 1.3 (IGD) out of 10 on synthetic tasks for self-IS ManGO (Zhou et al., 6 Jun 2025).

Another optimization-related use is MANGO: Learning Disentangled Image Transformation Manifolds with Grouped Operators, which replaces overlapping transport operators with grouped operators acting on distinct latent subspaces. The method enables one-phase training rather than MAE’s three-phase routine and reports a 100x speedup in the abstract, with runtime tables showing 319.50 for MAE versus 0.20 for MANGO at latent dimension $L=128$ in the reported setup (Ancelin et al., 2024).

Across these optimization-oriented uses, MANGO typically denotes a mechanism for structured adaptation under constraints: adaptive gradient control in continual learning, layered abstraction in HRL, joint manifold modeling in offline optimization, or grouped operator structure in latent transformation learning. This suggests that, despite the acronym’s semantic fragmentation, many MANGO systems are defined by explicit intermediate structure rather than by end-to-end undifferentiated optimization (Awasthi et al., 18 May 2026, Arcudi et al., 25 Aug 2025, Zhou et al., 6 Jun 2025, Ancelin et al., 2024).