Unmanned aerial vehicle (UAV) forest fire detection is vital for forest safety. However, early-stage UAV fire scenarios often involve small targets, weak smoke signals, and strict onboard resource constraints, which pose significant challenges to existing detectors. To improve the speed and accuracy of UAV forest fire detection, this paper proposes a lightweight fire detection algorithm, AHE-YOLO, specifically designed for UAVs. The proposed method adopts a coordinated lightweight design to improve feature preservation and cross-scale representation under limited computational budgets. Specifically, the Adaptive Downsampling (ADown) module preserves shallow fire-related cues during spatial reduction, improving sensitivity to small flame and smoke targets. The high-level screening-feature fusion pyramid network (HS-FPN) introduces cross-scale attention to promote more discriminative multi-level feature interaction while reducing redundant computation. Furthermore, the Efficient Mobile Inverted Bottleneck Convolution (EMBC) module is employed to improve receptive-field efficiency and feature selectivity under lightweight constraints, further enhancing detection accuracy and inference speed. Finally, the performance of AHE-YOLO is comprehensively evaluated through ablation and comparative experiments on the same dataset. The final experimental results show that YOLO-AHE achieves a mean average precision (mAP) of 94.8% while reducing model parameters by 39.7%, decreasing FLOPs by 27.0%, and shrinking the model size by 36.4%. In addition, its inference speed improves by 16.5%. Beyond detection performance, the proposed framework supports sustainable forest monitoring by enabling early fire warning with reduced computational and energy demands, showing strong potential for real-time deployment on resource-constrained UAV and edge platforms.

Abstract Accurate parameter identification is essential for predictive proton-exchange membrane fuel-cell (PEMFC) modelling, yet it remains challenging because commonly used semi-empirical formulations are strongly nonlinear and exhibit pronounced parameter coupling. This paper focuses on the optimisation engine rather than proposing a new PEMFC model. We introduce a sequential hybrid metaheuristic that combines the leader-guided exploration of the Grey Wolf Optimizer (GWO) with the adaptive exploitation mechanisms of Harris Hawks Optimization (HHO). The resulting method, denoted GWO--HHO, estimates the key parameters of a standard single-cell/stack model by minimising the sum of squared errors (SSE) between measured and simulated stack voltages. A unified identification protocol is applied to three public PEMFC benchmarks (Horizon 500~W, BCS 500~W, and Nedstack PS6), as well as to a suite of standard unimodal and multimodal test functions used to stress the optimiser itself. On the fuel-cell datasets, the proposed hybrid achieves SSE values of $0.0110$, $0.0116$, and $2.0655$, respectively, and exhibits smooth convergence with low variance across repeated runs. Compared with several recent optimisers under the same computational budget, GWO--HHO delivers competitive accuracy and robust behaviour across datasets. The method therefore provides a practical optimisation tool for reliable PEMFC parameter estimation using established physics-based models.

Resource-constrained intelligent systems increasingly require reliable long-form reasoning capabilities under limited computational and memory budgets, particularly in edge and embedded sensing environments. However, reinforcement learning for long-horizon decision generation remains highly unstable in such low-resource settings due to severe reward sparsity and imbalanced credit assignment, which often lead to non-convergent or excessively verbose generation behavior. In this work, we propose AF-CuRL (Answer-Focused Curriculum Reinforcement Learning), a lightweight reinforcement learning framework designed to stabilize long-form generation without increasing model size or computational cost. AF-CuRL improves optimization learnability through two complementary objective-level designs: (1) answer-focused token reweighting, which concentrates policy updates on reward-critical regions of generated sequences to alleviate credit assignment imbalance, and (2) a two-phase curriculum reward schedule that prioritizes stable termination and output regularity before shifting toward correctness-oriented optimization. We evaluate AF-CuRL on a 1.5B-parameter language model under strictly constrained training settings, using mathematical reasoning tasks as a controlled and reproducible proxy for long-horizon, rule-based decision-making commonly encountered in intelligent sensing and embedded systems. Experimental results demonstrate consistent improvements in both decision accuracy and generation regularity, including higher termination reliability and reduced generation length, compared with standard sequence-level reinforcement learning baselines. These results suggest that, for resource-limited and edge-intelligent systems, structured objective design can be more effective than model scaling for achieving stable and efficient long-form reasoning, providing a practical reinforcement learning solution for intelligent systems operating under real-world constraints.

ABSTRACT Recent grammar error correction (GEC) systems have scaled rapidly in model size and architectural depth, creating a growing mismatch between algorithmic improvements and the latency and energy constraints of edge devices. The method reformulates English GEC as a task‐constrained latent editing problem, where grammatical corrections are represented as low‐rank perturbations in a compact linear subspace. A Tiny‐LM‐style weight re‐parameterization aligns the latent editing vectors with a minimal set of re‐parameterized weights, ensuring that English grammatical reasoning is concentrated in a hardware‐friendly linear manifold. To improve correction fidelity under tight computational budgets, a two‐stage progressive refinement strategy is employed: a fixed‐window lookahead performs coarse structural edits, followed by a sparse consistency filter that selectively verifies candidate token corrections under INT8/INT4 quantization. The entire pipeline is static‐shape and operator‐regular, relying solely on linear, NPU‐native operations for predictable latency and bounded memory footprint. Experiments on public datasets show that the proposed model outperforms large Transformer baselines in F0.5 score on typical edge NPUs while reducing latency by 3–7×, demonstrating that accurate, low‐latency, on‐device English GEC is achievable using generic NPU operators without heavyweight language models.

In certain real-world optimization scenarios, practitioners are not interested in solving multiple problems but rather in finding the best solution to a single, specific problem. When the computational budget is large relative to the cost of evaluating a candidate solution, multiple heuristic alternatives can be tried to solve the same given problem, each possibly with a different algorithm, parameter configuration, initialization, or stopping criterion. In this practically relevant setting, the sequential selection of which alternative to try next is crucial for efficiently identifying the best possible solution across multiple attempts. However, suitable sequential alternative selection strategies have traditionally been studied separately across different research topics and have not been the exclusive focus of any existing review. As a result, the state-of-the-art remains fragmented for practitioners interested in this setting, with surveys either covering only subsets of relevant strategies or including approaches that rely on assumptions that are not feasible for the single-problem case. This work addresses the identified gap by providing a focused review of single-problem multi-attempt heuristic optimization. It brings together suitable strategies for this setting that have been studied separately through algorithm selection, parameter tuning, multi-start, and resource allocation. These strategies are described using a unified terminology within a common framework, which supports the construction of a taxonomy for systematically organizing and classifying them. The resulting comprehensive review facilitates both the identification and the development of strategies for the single-problem multi-attempt setting in practice.

Abstract Machine-learned interatomic potentials can offer near first-principles accuracy but are computationally expensive, limiting their application to large-scale molecular dynamics simulations. Inspired by quantum mechanics/molecular mechanics methods, we present ML-MIX, a CPU- and GPU-compatible package to accelerate simulations by spatially mixing interatomic potentials of different complexities, allowing deployment of modern MLIPs even under restricted computational budgets. We demonstrate our method for ACE, UF3, SNAP and MACE potential architectures and demonstrate how linear ‘cheap’ potentials can be distilled from a given ‘expensive’ potential, allowing close matching in relevant regions of configuration space. The functionality of ML-MIX is demonstrated through tests on point defects in Si, Fe and W-He, in which speedups of up to 11× over ~8000 atoms are demonstrated, without sacrificing accuracy. The scientific potential of ML-MIX is demonstrated via two case studies in W, measuring the mobility of $$b=\frac{1}{2}\langle 111\rangle$$ b = 1 2 〈 111 〉 screw dislocations with ACE/ACE mixing and the implantation of He with MACE/SNAP mixing. The latter returns He reflection coefficients which (for the first time) match experimental observations up to an He incident energy of 80 eV—demonstrating the benefits of deploying state-of-the-art models on large, realistic systems.

As autonomous driving technologies continue to evolve, their real-world applications are expanding beyond structured urban environments. However, off-road autonomous driving remains a challenging problem due to the unstructured and unpredictable nature of such terrains. For an unmanned ground vehicle (UGV) to drive reliably in off-road environments, a path planning algorithm must not only generate smooth trajectories that avoid abrupt changes but also ensure drivability by adapting to irregular terrain features. In this paper, we propose a novel and efficient path planning algorithm tailored for off-road driving. Our method, called two-step MPC-PSO, combines Model Predictive Control (MPC) with Particle Swarm Optimization (PSO) to generate optimal paths within a limited computational budget. We also design a cost function that explicitly accounts for off-road conditions to enhance terrain adaptability. We validate our approach through experiments conducted at two off-road test sites. The results demonstrate that our method generates paths with low maximum curvature and sufficient path length, enabling smooth and continuous driving in unstructured environments.

Towards lightweight stress monitoring on biometric data for IoMT environments.

Cyberbullying and hate speech increasingly appear in multi-modal social media posts, where images and text are combined in diverse and fast changing ways across platforms. These posts differ in style, vocabulary and layout, and labeled data are sparse and noisy, which makes it difficult to train detectors that are both reliable and deployable under tight computational budgets. Many high performing systems rely on large vision language backbones, full parameter fine tuning, online retrieval or model ensembles, which raises training and inference costs. We present a parameter efficient cross-platform multi-modal transfer learning framework for cyberbullying and hateful content detection. Our framework has three components. First, we perform domain adaptive pretraining of a compact ViLT backbone on in domain image-text corpora. Second, we apply parameter efficient fine tuning that updates only bias terms, a small subset of LayerNorm parameters and the classification head, leaving the inference computation graph unchanged. Third, we use noise aware knowledge distillation from a stronger teacher built from pretrained text and CLIP based image-text encoders, where only high confidence, temperature scaled predictions are used as soft labels during training, and teacher models and any retrieval components are used only offline. We evaluate primarily on Hateful Memes and use IMDB as an auxiliary text only benchmark to show that the deployment aware PEFT + offline-KD recipe can still be applied when other modalities are unavailable. On Hateful Memes, our student updates only 0.11% of parameters and retain about 96% of the AUROC of full fine-tuning.

IntroductionWith the rise of deep learning, automated electrocardiogram (ECG) classification has made huge improvements. Computer Vision (CV) and Natural Language Processing (NLP) are drivers of the deep learning revolution. While scaling model size has proven critical for performance in CV and NLP, medical data like ECGs exhibit fundamentally different properties. Despite this, scaling laws from CV/NLP are often applied directly to ECG tasks, leading to architecturally suboptimal models and prompting practitioners to default to smaller, simpler architectures that are easier to optimize.PurposeIn this work, we present a systematic framework for deriving scaling laws and architectural design principles of Convolutional Neural Networks (CNNs) for ECG classification. The aim is first to provide a framework to obtain model scaling laws, second to provide general scaling laws for CNNs for ECG classification, and finally to provide different CNN architectures for different computational budgets.MethodsWe propose a data-driven framework to derive ECG-specific scaling laws. Starting from a network design space encompassing CNN principles (kernel size, residual connections, inverted bottlenecks, depthwise convolutional blocks, squeeze-and-excitation blocks), we train 500 randomly sampled architectures. By analyzing performance distributions via Shapley additive explanations (SHAP values), we quantify the impact of each design choice and deduce optimal scaling laws.We used an open source dataset consisting of 88,000 twelve-lead ECG recordings from six different sources. The ECG classification task contains 30 labels describing cardiac abnormalities and sinus rhythm.ResultsOur framework provides three key insights for CNN architectures in ECG classification: First, systematic scaling of model depth and width contributes more significantly to performance than other architectural choices (Figure 1). Second, specific design elements - inverted bottlenecks and squeeze-and-excitation blocks - consistently improve classification accuracy. Third, across various computational budgets, the proposed models outperform state-of-the-art CNNs (Figure 2).ConclusionOur results demonstrate that architectural scaling laws tailored to ECG classification lead to more effective and efficient CNNs than direct adaptations of CV models. This marks a paradigm shift from blindly adapting CV architectures to embracing a data-driven, ECG-specific design approach that systematically uncovers optimal model configurations for computational budget constraints. While our framework is validated on a large multi-source dataset, its generalizability to rare arrhythmias or low-data settings remains to be tested. Additionally, future work could evaluate hybrid or attention-based models (e.g., Transformers) to further advance ECG classification.SHAP values for model performanceComparison to state-of-the-art CNNs

With the functions of egocentric observation and multimodal perception equipped in augmented reality (AR) devices, the next generation of smart assistants has the potential to reduce human labor and enhance execution efficiency in assembly tasks. Among diverse assembly activity understanding tasks, anticipating the near future activities is crucial yet challenging, which can assist humans or agents to actively plan and engage in interactions with the environment. However, the existing egocentric activity anticipation methods still struggle to achieve a decent trade-off between accuracy and computational efficiency, hindering them to be deployed in practical applications. To address this dilemma, in this paper, we propose a goal-guided prompting framework with adaptive modality selection (GP-AMS), for assembly activity anticipation in egocentric videos. For bridging the semantic gap between the historical observations and unobserved future activities, we inject the inferred high-level goal clues into the constructed prompts, which are further utilized to guide a pre-trained vision-language (V-L) model to compensate relevant semantics of unseen future. Moreover, a mask-and-predict strategy is adopted with two imposed constraints, i.e., casual masking and probabilistic token-dropping, to mine the intrinsic associations between the assembly activities within a specific procedure. For maintaining the benefits of exploiting multimodal information while avoiding extensively increasing the computational burdens, an adaptive modality selection strategy is designed to train a policy network, which learns to dynamically decide which modalities should be sampled for processing by the anticipation model on a per observation time-step basis. By allocating major computation to the selected indicative modalities on-the-fly, the efficiency of the overall model can be improved, thus paving the way for feasibility on real-world devices. Extensive experimental results on two public data sets validate that the proposed method yields not only consistent improvements in anticipation accuracy, but also significant savings in computation budgets.

Real-time global illumination rendered with low variance remains a persistent challenge. Many engines employ irradiance probes as a relatively cheap technique, but constrained computational budgets often lead to flickering artifacts in rendered images. In this paper, we propose spatial multiple importance sampling, which reuses ray-surface intersection data among real-time irradiance probes to significantly reduce flicker caused by variance, enabling efficient computation in complex scenes under limited ray tracing budgets. Moreover, our approach incorporates a probe selection mechanism to enhance reuse efficiency and a visibility estimation method to mitigate bias. Experimental results demonstrate that our method significantly reduces variance at a fixed ray tracing cost, delivering high-quality, stable outputs in real-time scenarios.

The integration of Battery Energy Storage Systems (BESS) is critical for mitigating voltage instability in low-voltage (LV) networks with high photovoltaic (PV) penetration. While metaheuristic algorithms offer powerful tools for optimizing BESS dispatch, their successful transition from theoretical models to practical application hinges on a nuanced understanding of their operational parameters. This paper presents a case study on the pragmatic design and tuning of a hybrid Particle Swarm Optimization-Grey Wolf Optimizer (PSO-GWO) for a six-dimensional, multi-BESS control problem. We chronicle the evolution of the simulation framework, highlighting critical implementation challenges and their solutions. Key findings demonstrate that optimizer population size, not just iteration count, is a decisive factor in control stability, particularly for computationally inexpensive configurations. We introduce a control oscillation metric as a key performance indicator and discuss the indispensable role of smart warm-starts and rate-limiting in generating physically viable and asset-safe control actions. The paper concludes that a successful BESS control strategy is defined not only by its ability to meet primary objectives like voltage regulation but also by the stability and practicality of the control signals it produces, presenting a crucial trade-off between computational budget and real-world viability.

Background. Continuous behavioral authentication (keystroke dynamics, touch/swipe, motion sensors) verifies identity without extra actions. However, models degrade under device, session and activity shifts, are sensitive to noise and often require significant labeling. As passwordless logins spread, demand rises for post-login risk control and for models that are robust, compute-efficient and stable in the wild. Objective. To develop and empirically study a compact CNN-Transformer hybrid with lightweight self-supervised masked time-series autoencoding (MAE-style) for mobile behavioral biometrics on the HMOG and WISDM datasets. Methods. A 1D-CNN front end extracts local cues from smartphone motion signals, while a Transformer encoder captures longer-range dependencies. We use masked reconstruction on unlabeled HMOG sessions for self-supervised pretraining under a limited computational budget and then fine-tune the same hybrid architecture for user identification. We evaluate three hybrid variants on HMOG (trained from scratch, with masked pretraining, and with masked pretraining plus CORAL domain adaptation) and three models on WISDM (a Transformer baseline, a hybrid trained from scratch and a hybrid initialized from the HMOG-pretrained weights). Performance is measured using user-level mean and median Equal Error Rate (EER) and AUC. Results. On HMOG, the hybrid model trained from scratch achieves the best user-level metrics (EER 21.51% mean, 18.63% median; AUC 0.854 mean, 0.905 median), while the lightweight MAE and CORAL variants do not yet surpass this baseline. On WISDM, the hybrid model substantially outperforms a pure Transformer baseline (EER 9.41% vs 51.25% mean; AUC 0.902 vs 0.488 mean), and cross-dataset initialization from the HMOG MAE-pretrained weights provides an additional improvement (EER 8.42% mean, 2.07% median; AUC 0.907 mean, 0.959 median). Conclusions. The results indicate that a compact CNN-Transformer hybrid is effective for sensor-based mobile behavioral biometrics and that even lightweight masked pretraining can be helpful for cross-dataset transfer. At the same time, the benefits of MAE and CORAL on HMOG depend strongly on the pretraining budget and masking configuration, suggesting that further tuning is needed to fully exploit self-supervised pretraining in this setting.

Metaheuristics are a fundament pillar of Industry 4.0, as they allow for complex optimization problems to be solved by finding good solutions in a reasonable amount of computational time. One category of important problems in modern industry is that of binary problems, where decision variables can take values of zero or one. In this work, we propose a binary version of the Pufferfish optimization algorithm (BPOA), which was originally created to solve continuous problems. The binary mapping follows a two-step technique, first transforming using transfer functions and then discretizing using binarization rules. We study representative pairings of transfer functions and binarization rules, comparing our algorithm with Particle Swarm Optimization, Secretary Bird Optimization Algorithm, and Arithmetic Optimization Algorithm with identical computational budgets. To validate its correct functioning, we solved binary problems present in industry, such as the Set Covering Problem together with its Unicost variant, as well as the Knapsack Problem. The results we achieved with regard to these problems were promising and statistically validated. The tests performed on the executions indicate that many pair differences are not statistically significant when both methods are already close to the optimal level, and significance arises precisely where the descriptive gaps widen, underscoring that transfer-rule pairing is the main performance factor. BPOA is a competitive and flexible framework whose effectiveness is mainly governed by the discretization design.

Lightweight vision models are widely deployed in mobile and embedded systems, where strict computational and memory budgets demand compact architectures. However, their evaluation remains dominated by ImageNet—a single, large natural-image dataset that requires substantial training resources. This creates a dilemma: lightweight models trained on ImageNet often reach capacity limits due to their constrained size, while scaling them to billions of parameters with specialized training tricks to achieve top-tier ImageNet accuracy does not guarantee proportional performance once the architectures are scaled back down to meet mobile constraints, particularly when re-evaluated on diverse data domains. These challenges raise two key questions: How should cross-dataset robustness be quantified in a simple and lightweight way, and which architectural elements consistently support generalization under tight resource constraints? To answer them, we introduce the Cross-Dataset Score (xScore), a simple metric that captures both average accuracy across domains and the stability of model rankings. Evaluating 11 representative lightweight models (2.5 M parameters) across seven datasets, we find that (1) ImageNet accuracy is a weak proxy for cross-domain performance, (2) xScore provides a simple and interpretable robustness metric, and (3) high-xScore models reveal architectural patterns linked to stronger generalization. Finally, the architectural insights and evaluation framework presented here provide practical guidance for measuring the xScore of future lightweight models.

Object detection plays a pivotal role in safety-critical applications, including autonomous driving, intelligent surveillance, and unmanned aerial systems. However, many state-of-the-art detectors remain highly resource-intensive; their large parameter sizes and substantial floating-point operations make it difficult to balance accuracy and efficiency, particularly under constrained computational budgets. To mitigate this accuracy-efficiency trade-off, we propose FE-YOLOv8, a lightweight yet more effective variant of YOLOv8 (You Only Look Once version 8). Specifically, two architectural refinements are introduced: (1) C2f-Faster (Cross-Stage-Partial 2-Conv Faster Block) modules embedded in both the backbone and neck, where PConv (partial convolution) prunes redundant computations without diminishing representational capacity; and (2) an EfficientHead detection head that integrates EMSConv (Efficient Multi-Scale Convolution) to enhance multi-scale feature fusion while simplifying the head design and maintaining low computational complexity. Extensive ablation and comparative experiments on the SODA-10M dataset show that FE-YOLOv8 reduces the parameter count by 31.09% and the computational cost by 43.31% relative to baseline YOLOv8 while achieving comparable or superior mean Average Precision (mAP). Generalization experiments conducted on the BDD100K dataset further validate these improvements, demonstrating that FE-YOLOv8 achieves a favorable balance between accuracy and efficiency within the YOLOv8 family and provides new architectural insights for lightweight object detector design.

Early identification of mental health risks among college students is critical for timely intervention, promoting well-being, and supporting academic performance. This study utilizes a comprehensive multimodal dataset comprising 1,000 students, integrating behavioral routines (study hours, sleep schedules, and daily activity patterns), physiological indicators (heart rate, stress levels, and sleep quality), and social engagement measures (messaging frequency and participation in clubs or events) to classify students into Low, Moderate, and High mental health risk categories. Data preprocessing included handling missing values with mean/median imputation for continuous features and mode imputation for categorical features, followed by standardization using Z-score normalization. Stratified five-fold cross-validation with a fixed random seed was applied to ensure reproducible and unbiased evaluation. Baseline models, including the Data Fusion Model, CASTLE, YOLOv8, Time-Aware Multimodal Fusion Network (TAMFN), and Random Forest combined with CatBoost, were carefully tuned under equivalent computational budgets to provide fair comparisons. The proposed Augmented Wingsuit–Enhanced Multinomial Naïve Bayes (AWNB) framework combines optimization-driven hyperparameter tuning with decision-level multimodal fusion, effectively capturing complex interactions between behavioral, physiological, and social features. Experimental results demonstrate that AWNB achieves superior performance, with 97.41% accuracy, 95.14% precision, 93.67% recall, and 94.82% F1-score. Baseline performances were: Data Fusion Model – 95.2% accuracy, 93.7% precision, 90.8% recall, 92.2% F1-score; CASTLE – 84.47% accuracy, 71.47% recall, 74.65% F1-score; YOLOv8 – 71% precision, 74.1% recall; TAMFN – 66.02% precision, 66.50% recall, 65.82% F1-score; and Random Forest + CatBoost – 91.3% accuracy, 92.4% precision, 90.5% recall. All metrics are reported as mean ± standard deviation, and statistical significance was validated using paired tests. These findings establish AWNB as a robust, interpretable, and computationally efficient framework, outperforming existing approaches while enabling scalable application in academic mental health monitoring.

While real-time multi-person pose estimation is a critical technology for human-computer interaction and action recognition tasks, maintaining accuracy and efficiency on confined hardware remains a major challenge. To overcome the inherent trade-off between the high computational cost of heatmap-based methods and the inferior quality of regression-based ones, this paper uses a coarse-to-fine deep learning mechanism to propose a novel two-stage model named Dynamic Resolution Pose Network (DRP-Net). The model employs a light regression head first for rapid coarse coordinate estimation, then a dynamic refinement head to produce localized heatmaps in small, dense regions of interest to enable precise correction. This effectively maximizes the utilization of computation resources and provides high localization accuracy with significantly reduced model inference latency. Experimental results verify that the medium-sized DRP-Net-M model achieves an Average Precision (AP) of 74.1% on the MS COCO test set at a computation cost of mere 2.15 GFLOPs, outperforming the best-performing real-time model RTMPose-m with a comparable computational budget. This paper presents a two-stage architecture integrating regression and region-localized heatmap refinement. It provides a new high-efficiency paradigm for light-weight real-time pose estimation and sets a new direction to build other dense prediction tasks in computer vision through its dynamic resolution concept.

Text-to-3D scene generation is pivotal for digital content creation; however, existing methods often struggle with global consistency across views. We present 3DS-Gen, a modular “generate-then-reconstruct” framework that first produces a temporally coherent multi-view video prior and then reconstructs consistent 3D scenes using sparse geometry estimation and Gaussian optimization. A cascaded variational autoencoder (2D for spatial compression and 3D for temporal compression) provides a compact and coherent latent sequence that facilitates robust reconstruction. An adaptive density threshold improves detailed allocation in the Gaussian stage under a fixed computational budget. While explicit meshes can be extracted from the optimized representation when needed, our claims emphasize multiview consistency and reconstructability; the mesh quality depends on the video prior and the chosen explicitification backend. 3DS-Gen runs on a single GPU and yields coherent scene reconstructions across diverse prompts, thereby providing a practical bridge between text and 3D content creation.