Paper deep dive
A Multi-Modal Generative Model for Tomato Disease Leaves Understanding
Khang Nguyen Quoc, Minh-Phuoc Tran, Gia-Han Truong, Luyl-Da Quach
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 9/18/2026, 4:24:19 AM
Summary
The paper introduces SOLAR, a multimodal generative model designed for tomato disease leaf understanding. It reformulates disease analysis as a Visual Question Answering (VQA) task to integrate visual symptoms with semantic context. SOLAR utilizes a Fusion Expert module based on Mixture-of-Experts (MoE) to align visual features with task-aware language representations, enabling it to perform six diagnostic tasks across three hierarchical levels: Basic Perception, Pathology Understanding, and Disease Diagnosis. The model was evaluated on 41,677 images and 216,209 QA pairs, demonstrating superior performance over state-of-the-art vision-only and vision-language models.
Entities (16)
Relation Signals (13)
SOLAR → appliesto → Tomato Disease
confidence 98% · Here, we present SOLAR, a multimodal generative model that understands tomato disease
SOLAR → solvestask → Visual Question Answering
confidence 95% · By formulating tomato disease analysis as a generative Visual Question Answering (VQA) task, SOLAR provides a flexible framework
SOLAR → usescomponent → Fusion Expert Module
confidence 95% · SOLAR learns to align visual features with task-aware language representations by Fusion Expert module based on mixture-of-expert
Fusion Expert Module → basedon → Mixture-of-Experts
confidence 92% · Fusion Expert module based on mixture-of-expert
SOLAR → outperforms → state-of-the-art vision-only, vision-language, and task-specific models
confidence 92% · Experimental results show that SOLAR consistently outperforms state-of-the-art vision-only, vision-language, and task-specific models across all tasks
SOLAR → performstask → Scientific Name Classification
confidence 90% · Scientific Name Classification (SNC) tasks present the most demanding challenge, requiring the model to identify the scientific nomenclature
SOLAR → performstask → Leaf Count
confidence 90% · Leaf Count (LC) tasks require the model to identify and enumerate the leaves present in a given image
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Artificial intelligence for plant disease analysis has advanced from task-specific classifiers to multi-modal models capable of jointly interpreting visual and textual information. However, practical deployment in precision agriculture remains limited because most existing approaches treat disease understanding as isolated prediction tasks, failing to capture the complementary relationships among symptom recognition, severity assessment, and question-driven diagnostic reasoning. In tomato pathology, accurate interpretation of diseased leaves requires more than label prediction; it demands integrating visual symptoms with semantic context to support a comprehensive and explainable understanding. Here, we present SOLAR, a multimodal generative model that understands tomato disease spanning six question-answering tasks. SOLAR learns to align visual features with task-aware language representations by Fusion Expert module based on mixture-of-expert, enabling it to generate contextually relevant answers across diverse diagnostic tasks. By formulating tomato disease analysis as a generative Visual Question Answering (VQA) task, SOLAR provides a flexible framework that supports multi-task inference within a single model while improving performance and cross-task knowledge sharing. We evaluate SOLAR on $41,677$ images, including $216,209$ Question-Answering (QA) pairs to understand tomato leaf disease under both closed and open-ended QA settings. Experimental results show that SOLAR consistently outperforms state-of-the-art vision-only, vision-language, and task-specific models across all tasks, demonstrating superior accuracy, robustness, and multimodal reasoning. These findings highlight the potential of generative multimodal modeling as an effective direction for understanding of plant disease. The code for this study is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2609.19555v1
- Canonical: https://arxiv.org/abs/2609.19555v1
Trouble viewing inline? Open PDF directly →
Full Text
113,553 characters extracted from source content.
Expand or collapse full text
A Multi-Modal Generative Model for Tomato Disease Leaves Understanding Journal: Computers and Electronics in Agriculture Khang Nguyen Quoc Email: khangnq@korea.ac.kr Affiliation: School of Electrical Engineering, Korea University, Seoul, 02841, South Korea Minh-Phuoc Tran Email: minhphuoc19122005@gmail.com Affiliation: Department of Software Engineering, FPT University, Cantho city, CT 90000, Vietnam Gia-Han Truong Email: truonggiahan0702ct@gmail.com Affiliation: Department of Software Engineering, FPT University, Cantho city, CT 90000, Vietnam Luyl-Da Quach Email: luyldaquach@gmail.com Affiliation: Department of Software Engineering, FPT University, Cantho city, CT 90000, Vietnam Corresponding author: Corresponding author Abstract Artificial intelligence for plant disease analysis has advanced from task-specific classifiers to multi-modal models capable of jointly interpreting visual and textual information. However, practical deployment in precision agriculture remains limited because most existing approaches treat disease understanding as isolated prediction tasks, failing to capture the complementary relationships among symptom recognition, severity assessment, and question-driven diagnostic reasoning. In tomato pathology, accurate interpretation of diseased leaves requires more than label prediction; it demands integrating visual symptoms with semantic context to support a comprehensive and explainable understanding. Here, we present SOLAR, a multimodal generative model that understands tomato disease spanning six question-answering tasks. SOLAR learns to align visual features with task-aware language representations by Fusion Expert module based on mixture-of-expert, enabling it to generate contextually relevant answers across diverse diagnostic tasks. By formulating tomato disease analysis as a generative Visual Question Answering (VQA) task, SOLAR provides a flexible framework that supports multi-task inference within a single model while improving performance and cross-task knowledge sharing. We evaluate SOLAR on 41,67741,677 images, including 216,209216,209 Question-Answering (QA) pairs to understand tomato leaf disease under both closed and open-ended QA settings. Experimental results show that SOLAR consistently outperforms state-of-the-art vision-only, vision-language, and task-specific models across all tasks, demonstrating superior accuracy, robustness, and multimodal reasoning. These findings highlight the potential of generative multimodal modeling as an effective direction for understanding of plant disease. The code for this study is available at https://github.com/EnalisUs/SOLAR. Keywords: Visual Question Answering , Multi-modal models , Tomato Disease , Vision-language model , large language models †graphicalabstract: Overview of the SOLAR framework for tomato leaf disease diagnosis via Visual Question Answering (VQA). The architecture comprises two input modalities: a Vision branch, in which input leaf images are processed through a Vision Encoder and subsequently refined by an Expert Fusion module consisting of a Router, multiple specialized Experts, and a Projector that maps visual features into the language embedding space; and a Language branch, in which natural language queries are tokenized and encoded as text embeddings. Both modalities are jointly processed by LLM to produce responses across three hierarchical diagnostic levels–Basic Perception, Pathology Understanding, and Disease Diagnosis—enabling structured, coarse-to-fine disease recognition across six VQA task types. †highlights: This study is the first to reformulate the tomato image classification problem as a generative task, specifically, Visual Question Answering. We propose SOLAR, a generative multimodal model for understanding tomato disease that mimics an agricultural expert’s diagnostic workflow across multiple VQA tasks. We propose an Expert Fusion module to improve alignment between the visual feature space and multiple question embeddings. We evaluate SOLAR on a combined set of four distinct tomato disease datasets, which we formulate into six diagnostic tasks. Our model achieves state-of-the-art (SOTA) performance across all benchmarks, outperforming existing SOTA vision, vision-language, general, and domain-specific foundation models. 1 Introduction According to statistics from the European Commission, tomato production reached over 8.2 million tons (2025). However, it showed a downward trend from 2024 (over 8.6 million tons), indicating that tomatoes are among the most important crops globally [10]. This production tends to decrease due to strong seasonal and disease-related influences. To tackle this problem, many studies have focused on applying artificial intelligence (AI) and advances in deep learning (DL) to classify and name diseases, but have not yet optimized feature classification and the embedding space of large language models (LLMs) [4, 51]. Therefore, in this problem, the research focuses on purely visual image classification in the form of VQA, with specific tasks, to serve as a foundation for subsequent studies that build datasets and models for each specific disease. In AI applications for smart agriculture, research often focuses solely on disease identification, severity assessment, and symptom localization as separate tasks. In the disease identification task, research surveys leaf image classification approaches that leverage single-leaf characteristics using traditional machine learning and DL models. However, these can only answer "what disease is this leaf?" and fail to generalize to unseen content [17]. Meanwhile, disease detection approaches also identify diseased areas using bounding boxes with Faster R-CNN [31], YOLO [30], and related models. While development has shifted from two-stage to single-stage models, blurred and irregular symptom borders affect severity assessment and make it difficult to identify small lesions, such as early-stage spots [28]. Furthermore, semantic or instance segmentation techniques for symptom localization, severity measurement, and disease monitoring are expensive, require expert support, and do not provide sufficient information for disease identification [53]. These challenges make it difficult to detect early abnormal symptoms in plants, compounded by semantic correlations between images and disease symptoms in the complex environment of smart agriculture and the level of farmers’ understanding. The problem of disease detection in tomato plants is no exception to this trend; currently, the focus is only on detecting, classifying, and isolating diseases, and there are common limitations in smart agriculture. Research focuses on detecting and recognizing small objects in complex real-world environments by moving from feature engineering to semantic representation to identify meaningful regions, making it easier to access new data in these scenarios, but it does not yet understand the causal relationships and disease progression [48]. Research using prescription data combined with stacking ensemble learning and an image semantics approach differs in that it shifts from pixel-level to context-level semantics, which is less dependent on images, but it does not directly address the full semantic understanding problem in computer vision, as it does not truly address the problem when combining multimodal fusion [51]. The research shows improvements in data enhancement and optimization for disease, pest, and weed detection, with the potential to enhance the semantic representation of small, obscured objects by leveraging attention to increase semantic focus in YOLO. However, it does not yet understand the depth level [25]. This indicates that traditional classifier models act like a “black box,” only providing labels without explaining why, making it difficult for agricultural experts to trust them. To address the limitations identified in previous studies, we introduce SOLAR, a generative multimodal model. SOLAR innovates by framing the problem as a VQA task informed by the diagnostic process of task-specific agricultural experts, and implemented across six integrated diagnosis tasks. At the same time, SOLAR improves the semantic relationship between images and text through proposed modules and evaluations based on different datasets. The main contributions of this work are summarized as follows: 1. This study is the first to reformulate the tomato image classification problem as a generative task, specifically, Visual Question Answering. 2. We propose SOLAR, a generative multimodal model for understanding tomato disease, capable of mimicking an agricultural expert’s diagnostic workflow across multiple VQA tasks. 3. We propose an Expert Fusion module to improve alignment between the visual feature space and multiple question embeddings. 4. We evaluate SOLAR on a combined set of five distinct tomato disease datasets, which we formulate into six diagnostic tasks. 5. Our model achieves SOTA performance across all benchmarks, outperforming existing SOTA vision, vision-language, general, and domain-specific foundation models. 2 Related Work 2.1 Deep Learning in Plant Disease Diagnosis The DL method based on Convolutional Neural Networks (CNNs) and Vision Transformers has made progress in classifying various plant disease characteristics. In particular, the use of basic DL models such as ResNet and EfficientNet, along with advances in parameter optimization and adjustment, has achieved high accuracy in identifying and classifying leaf diseases in maize, rice, wheat, and mango, with accuracies exceeding 90%. However, it focuses only on classification and lacks explanatory indices for the results [35, 18, 8]. To improve model accuracy, many studies have combined advances in DL, such as transformer-based architectures and attention mechanisms, into YOLOv4 to delineate disease regions on leaf backgrounds and in the medium [48, 25]. In other studies, utilizing the advancements of LLM models has made certain progress in interpreting results through visualizing features with color (Grab-CAM, Ablation-CAM), or using text to clarify features, as studies conducted on corn, tomatoes, and mangoes have achieved positive results, as it evaluates the relationship between features related to disease characteristics, physiology and contributes to expressing semantic information at the expert level but has not yet expressed the features or simulated the expert’s diagnostic process [33, 50, 34]. These issues challenge the ability to leverage advances in computer vision and natural language processing models to interpret and infer, thereby providing an expert perspective rather than analyzing images pixel-by-pixel or simply reading plain text. 2.2 Vision-Language Models in Agriculture Addressing the limitations of traditional image classification, the VLM model is proposed to shift data from "image-only" to "image-contextual text" in agriculture. Various approaches are employed to enhance the effectiveness of Vision-Language Models (VLMs) using generative adversarial networks (GANs). In the applied approach, research focuses on proposing a pragmatic framework and exploring statistical language models, neural network models, and big language models, intending to integrate agricultural knowledge into VLMs using approaches such as instruction tuning, in-context learning, and generative retrieval architectures. However, these approaches also face difficulties in terminology and information discrepancies, such as distribution shift [23, 54]. Another approach involves building high-complexity standard datasets to test, evaluate, and identify the limits of existing VLMs across various tasks, such as AgroCoT (5 tasks) and AgEval (12 tasks), to assess in-context learning capabilities. However, these models are limited in accuracy, as evidenced by their complete reliance on GANs such as GPT, Gemini, and LlaVA, where performance decreases significantly when encountering types less common in the general dataset (AgEval) or when generating redundant text (AgroCoT) [49, 3, 36]. A completely different approach, studies propose new models capable of generating in-depth VLMs for agriculture based on different approaches (based on LLaVA/Mipha to combine with SigLIP/CLIP ViT with LLM or based on contrastive learning to combine image and text encoders into a common embedding space), which creates effective plant disease identification models in various scenarios, but is limited in its ability to infer open conversational patterns and suffers from logical errors when relying on GAN models [5, 37]. Overall, research studies have shown a shift away from rigid classification labels toward detailed symptom analysis, data smoothing, expert support, and step-by-step explanations. However, this also presents challenges due to the need for expert knowledge, which can be overcome by adopting the flexible, natural-language-based approach of generative VQA. 2.3 Generative VQA with interpretive capabilities The VQA model is an artificial intelligence system that takes images and natural-language questions as input, then provides accurate answers through three structured stages: a vision encoder, a language encoder, and a fusion & decoder. This helps to answer specific, clear questions, demonstrating true semantic and visual inference capabilities [2]. In the traditional VQA approach, the method involves selecting answers from a fixed list of labels or filling in the blanks through knowledge processing by directly embedding into network layers, such as co-attention and multimodal tucker fusion with free responses [52], combining CNN with transformer models to extract pathological image features and link information with questions from farmers [20], and combining VQA with federated learning models to update weights with recognition questions [27]. Its advantage is its ease of implementation on edge devices, but it cannot answer questions outside the predefined set of labels. In particular, VQA relies on a large-scale/dialogue model to facilitate free-text generation, answer open-ended questions, and provide highly detailed technical solutions through LLM reasoning, such as mapping 5 core information aspects to 5 core information aspects, but it fails to overcome complex situational reasoning questions [36, 32]. PlantVillageVQA applies a language reconstruction system and diverse natural-language questions with expert review, but it faces significant accuracy challenges when real-world images have complex backgrounds [39]. Research using authentic grower dialogues, through tests and open-ended questions, helps capture farmers’ writing styles and the real-world issues they care about [13]. Research using contrastive learning combined with lesion-aware learning and LLM models to conclude is limited to the sketch level[40]. However, these models suffer from technical and computationally expensive hallucinations. Overall, research shows that the VQA approach overcomes one-hot labels (e.g., forcing the model to respond as “disease” or “healthy”) by consolidating tasks through prompt changes, but still relies on question-embedding sensitivity and true interpretability. Hence, our study developed an expert fusion module to align the space and act as an expert to analyze the data and apply true pathology logic. 3 Method 3.1 Task Formulation To transcend the limitations of conventional single-label image classification, this study reformulates tomato leaf disease recognition as a structured VQA problem organized around a coarse-to-fine diagnostic reasoning hierarchy. Formally, given an input image I and a natural language question Q, the model ℱF is trained to generate a textual answer =f(I,)A=f(I,Q), where f denotes the vision-language model. Under this formulation, we define six VQA tasks grouped into three progressive levels of diagnostic reasoning, encompassing basic visual perception, pathology understanding, and disease diagnosis, following: 1. Level I: Basic Perception. The first group establishes foundational visual awareness prior to any disease-level reasoning. Leaf Count (LC) tasks require the model to identify and enumerate the leaves present in a given image, probing spatial and object-level perceptual capability. Healthy or Disease Classification (HDC) tasks subsequently demand a coarse binary judgment of whether the observed leaf is healthy or exhibiting visible signs of pathological change, serving as an initial diagnostic screening gate. 2. Level I: Pathology Understanding. Building upon perceptual foundations, the second group introduces a higher degree of biologically grounded reasoning. Pathogen Classification (PC) tasks require the model to infer the nature of the causative agent—distinguishing among fungal, bacterial, and viral origins—based solely on observable visual evidence. Symptom Identification (SI) tasks further demand that the model recognize and articulate specific visible disease manifestations, including lesions, chlorosis, necrosis, and irregular spotting patterns, thereby grounding its responses in directly observable leaf morphology. 3. Level I: Disease Diagnosis. The third group demands precise, clinically meaningful outputs that reflect deep domain knowledge. Disease Classification (DC) tasks require the model to assign the correct disease label from among 15 predefined tomato disease categories, necessitating fine-grained visual discrimination. Scientific Name Classification (SNC) tasks present the most demanding challenge, requiring the model to identify the scientific nomenclature of the diagnosed condition—a degree of pathological specificity that extends well beyond colloquial naming conventions. Collectively, these three hierarchical levels constitute a coherent diagnostic reasoning pipeline that progresses systematically from low-level visual perception, through biological pathology interpretation, to high-precision disease identification, mirroring the reasoning process of domain experts in plant pathology. 3.2 SOLAR Architecture SOLAR is a multimodal generative model designed to emulate the diagnostic reasoning workflow of an agricultural expert in the context of tomato leaf disease understanding. Given a leaf image as input, SOLAR first encodes the visual content into a rich set of feature representations that capture the biological characteristics of the observed leaf. A structured two-stage reasoning pipeline then processes these features to produce accurate, contextually grounded responses. In the first stage, expert-aware visual encoding, the extracted visual features are refined by a Mixture-of-Experts (MoE) [12] module that selectively activates relevant feature pathways conditioned on the input question. This mechanism allows the model to adaptively emphasize diagnostically relevant visual regions, such as lesion patterns, discoloration, or necrotic spots, depending on the nature of the query. The filtered visual features are subsequently fused with the question embedding to form robust multimodal representations, jointly conditioned on both the visual input and the question’s linguistic context. In the second stage, answer generation, a large language model (LLM) receives the question embedding alongside the expert-aware visual features to generate a free-form textual response tailored to the current task. This design enables SOLAR to flexibly handle the full spectrum of diagnostic queries across the defined coarse-to-fine task hierarchy, from basic leaf perception to detailed pathological diagnosis. By aggregating evidence across all reasoning stages, SOLAR produces a comprehensive and interpretable understanding of tomato leaf disease, bridging the gap between visual observation and expert-level diagnostic judgment. The framework consists of three primary components: 1) an image encoder fimgf_img, 2) an Expert Fusion module ℰE, and 3) a LLM decoder T. The overall architecture is illustrated in Figure 1. Figure 1: Overview of the SOLAR framework for tomato leaf disease diagnosis via VQA. The architecture comprises two input modalities: a Vision branch, in which input leaf images are processed through a Vision Encoder and subsequently refined by an Expert Fusion module consisting of a Router, multiple specialized Experts, and a Projector that maps visual features into the language embedding space; and a Language branch, in which natural language queries are tokenized and encoded as text embeddings. Both modalities are jointly processed by LLM to produce responses across three hierarchical diagnostic levels–Basic Perception, Pathology Understanding, and Disease Diagnosis–enabling structured, coarse-to-fine disease recognition across six VQA task types. Given a image ℐI, we extract visual embedding V by the visual encoder fimgf_img, where ∈ℝDV ^D. We utilize SCOLD [37], a vision-language foundation model pretrained on over 157K plant image-caption pairs, as the visual encoder. Since the dimension of the image encoder is different from the text embeddings, we use a projection layer to map P into the same space as the embeddings =()V=P(V). The question-aware Expert Fusion adapter ℰE bridges the visual embedding space and the LLM embedding space. Unlike traditional static projectors (e.g., LLaVA-OV [21]), which directly map visual embedding to LLM space, ℰE explicitly incorporates the question embedding Q as an expert filter, obtained by encoding the question with a text encoder, to produce task-relevant experts. To obtain Q, we leverage LFM-VL 2.5 [1], a model designed for on-device deployment. It builds on the LFM2 architecture with extended pre-training and reinforcement learning. Building on this encoder-decoder model, we ensure that the semantic feature space of the encoded question is well-designed for the generative space. With the features extracted from the question and the image, the Expert Adapter module ℰE first refines the visual features via a Sparse-MoE to filter the important information relevant to each specific question.A remarkable aspect of our model is that the gating network takes the question Q as input to determine which experts to activate. Consequently, the same image paired with different questions will trigger different experts, it make visual features for each question type. Ve=+∑i=1NRi()⋅Ei()V_e=V+ _i=1^NR_i(Q)· E_i(V) (1) With N is the number of experts, EiE_i denotes the i-th expert layer, and RiR_i is the i-th routing weight computed from the question Q from the gating network. Specifically, only the K experts with the highest routing weights are evaluated. Subsequently, the expert-aware feature fuses with the question feature to compute the final visual token with flexibility for each question. xv=Dropout(GELU(LayerNorm(Wf⋅[||Ve])))x_v=Dropout (GELU (LayerNorm(W_f·[Q||\,V_e]) ) ) (2) Where [⋅||⋅][·||·] represents the concatenation operation and WfW_f are learnable projection weights. The LLM decoder, denoted T, executes the VQA task by formulating diagnostic responses in natural language. In contrast to conventional classification models that only predict classes, T accepts the specific visual token XvX_v alongside the question tokens XqX_q. After concatenating these inputs, it generates the response autoregressively, predicting each subsequent token conditioned on both the visual features and the textual context. 3.3 Question-Answer Dataset Construction We develop the dataset in a question–answer prompt format, where each image is paired with a corresponding question and answer. For each type of disease, a sample consists of a leaf image, a diagnostic question, and its associated ground-truth label. These pairs are represented as (XQ,i,XA,i)i=1N(X_Q,i,X_A,i)_i=1^N where XQ,iX_Q,i denotes the input question related to the image, and XA,iX_A,i denotes the corresponding ground-truth output. This formulation enables the model to learn task-specific diagnostic knowledge, focus on visual information related to the question, follow a fixed output format, and generate correct labels. We construct the input sequence by concatenating the question XQX_Q, the string “ Answer: ”, and the ground-truth answer XAX_A to make a unified prompt: Xp=XQ+“ Answer: ”+XAX_p=X_Q+`` Answer: ′+X_A (3) 3.4 SOLAR Training and Inference The full SOLAR training target is presented below: L=LTASK+λ1LAUX+λ2LDIVL=L_TASK+ _1L_AUX+ _2L_DIV (4) where LTASKL_TASK is the loss of autoregressive language modeling, LAUXL_AUX is the loss of load balance that regularizes expert utilization, and LDIVL_DIV is the loss of expert diversity that encourages specialized representations. λ1 _1 and λ2 _2 denote weighting coefficients. The primary loss LTASKL_TASK maximizes the probability of the target answer XAX_A consisting of M tokens conditioned on the input prompt XpX_p, which is given by: LTASK=−∑i=1Mlogpθ(xi∣Xp,x<i)L_TASK=- _i=1^M p_θ(x_i X_p,x_<i) (5) where, θ denotes the trainable parameters, xix_i is the token at position i in the answer, and x<ix_<i refers to all tokens that appear before position i. In the Mixture-of-Experts architecture, a common issue is expert collapse, where the gating network repeatedly routes most inputs to only a few experts, while the remaining experts are barely used. To prevent this degenerate behavior, we apply the auxiliary load-balancing loss LAUXL_AUX following the formulation of the Switch Transformer [11]: ℒAUX=N⋅∑i=1Nfi⋅PiL_AUX=N· _i=1^Nf_i· P_i (6) where fif_i is the fraction of samples in batch B sent to expert i. fi=1B∑b∈ℬexpert i∈TopK(b)f_i= 1B _b 1\expert i (b)\ (7) and PiP i is the fraction of the probability that the router is assigned to the expert i. Pi=1B∑b∈ℬpi(b)P_i= 1B _b p_i(b) (8) with pi(b)p_i(b) being the probability of gating assigned to expert i in the sample b in the batch. This loss reaches its minimum when all experts are utilized uniformly, fi=Pi=1Nf_i=P_i= 1N for all i. Although load balancing helps distribute inputs relatively evenly between experts, it does not ensure that different experts learn different features. We observe that, without explicit regularization, experts may learn overlapping features, which reduces the effectiveness of the model. To address this issue, based on the Mean Squared Overlap [19], we propose an expert diversification loss LDIVL_DIV that penalizes high cosine similarity between the weight matrices of different experts. Let Wi∈ℝdh×dvW_i ^d_h× d_v denote the weight matrix of the first linear layer in expert i, where dhd_h is the hidden dimension and dvd_v is the visual dimension. We then flatten and normalize each weight matrix in L2 into a unit vector w^i w_i, and define: ℒDIV=1N(N−1)∑i<j(^i⊤^j)2L_DIV= 1N(N-1) _i<j( w_i w_j)^2 (9) This loss helps experts specialize better, ensuring that they extract different aspects of visual features. In the training phase, we fine-tune the model on the TomaMMU dataset [45], which includes more than 27,000 unique images and 94,932 QA pairs. These QA pairs are divided into six specific groups, covering three levels from simple to complex: Basic Perception, Pathology Understanding, and Disease Diagnosis. This design helps the model to understand the question more effectively. This allows us to pay attention to the right visual evidence in the image and make the best diagnoses. Furthermore, in the LLM layer, we apply LoRA[15] to optimize performance, and the learned weights are merged directly into the original LLM weights. Specifically, for a frozen weight matrix W0∈ℝd×kW_0 ^d× k of the attention layer, the weight update ΔW W is decomposed into the product of two low-rank matrices A∈ℝr×kA ^r× k and B∈ℝd×rB ^d× r with rank r≪min(d,k)r (d,k): W=W0+ΔW=W0+αrBAW=W_0+ W=W_0+ αrBA (10) The forward pass calculation for a given hidden activation vector h through the adapter-equipped attention layer is formulated as: h′=W0h+αrBAh =W_0h+ αrBAh (11) where α is a constant scaling hyperparameter. Throughout training, the original pre-trained weight matrix W0W_0 remains completely frozen, and gradients are only computed and backpropagated to update the low-rank adapter parameters A and B, which significantly minimizes memory consumption and training overhead. After training phase, we evaluate the model on the validation set at the end of each epoch. We select the model weights from the epoch that achieves the highest validation accuracy as the final checkpoint for testing phases. In the inference phases, the model uses the components described in the Architecture section. Given the input image I and the question Q, the question-conditioned visual token xvx_v is computed using the Expert Fusion adapter ℰE: xv=ℰ(,)x_v=E(V,Q) (12) Then, the target diagnostic answer Y=(y1,y2,…,yL)Y=(y_1,y_2,…,y_L) is generated in autoregressive form based on the visual token xvx_v and the question tokens, based on the following conditional probability equation P(Y∣xv,)=∏t=1LP(yt∣y<t,xv,)P(Y x_v,Q)= _t=1^LP(y_t y_<t,x_v,Q) (13) where y<ty_<t denotes the sequence of answer tokens generated prior to step t. 4 Experimental Setup 4.1 Datasets We evaluated SOLAR in our internal and three external tomato disease datasets, including TomaMMU [45], TLID [56], TLD-3 [41] and TomaAD [40]. All input images were resized to a fixed resolution of 224×224224× 224 pixels prior to model training and testing. The overview of the datasets is shown in Figure 2. Figure 2: Distribution of QA samples(a) and Number of classes(b) across the six VQA tasks for the five datasets. In TomaMMU, number of class SNC and DC are differ because two DC classes, Mite and Spider mites Two-spotted Spider Mite, are the same scientific name Tetranychus urticae. In the same way two DC classes, Mosaic Virus and Virosis, are same Tomato mosaic virus (ToMV) in SNC. In this study, we utilized our TomaMMU dataset [45]. The images were filtered according to disease, symptom, leaves, pathogen, scientific name, and image descriptions. Lastly, the research team classified the images into disease types. The Dataset consists of more than 21,000 images of 15 diseases with 6 tasks, 158,214 QA pairs, and covering 6060 classes. TomaMMU was divided into 6:2:2, which corresponds to 94,932 QA pairs in the training phase, 31,641 QA pairs in the validation phase, and 31,641 QA pairs in the testing phase. We ensure that the images belonging to each disease category are proportionally allocated within the dataset, despite the inherently unbalanced distribution between classes. The statistics of these dataset classes are presented in Figure 1. The three external test datasets are reserved solely for the testing phase. Before entering this phase, these images are filtered and preprocessed following TomaMMU [45] procedure. TLID [56] was constructed through multiple field surveys in various greenhouse farms, allowing the documentation of different diseases at various stages of development. This dataset comprises seven distinct classes categorized into 4 types of disease , two co-infection scenarios, and one healthy state. However, to align with our model, we utilized only 2 disease with 25,973 QA pairs. TLD-3 [41] is a dataset specifically focused on images of tomato leaf collected under various environmental conditions. This dataset comprises 10 classes with 49,213 QA pairs representing various leaf diseases. TomaAD [40] is a dataset consisting of 502 tomato leaf images categorized into 4 diseases with 3,174 QA pairs. Data were collected from a commercial greenhouse in the arid climate of Abu Dhabi, United Arab Emirates. Disease labels were assigned based on visual observation and validated by agricultural experts on-site. In particular, these datasets reflect real-world imaging conditions, including illumination variance, leaf orientation, and partial occlusions. 4.2 Comparison Setup State-of-the-art competing models. To benchmark SOLAR, we evaluated competing SOTA architectures across three distinct paradigms. First, to assess out-of-the-box generalization, we tested a cohort of FMs and domain-specific VLMs without task-specific fine-tuning, comprising InternVL3 (1B and 2B) [55], Qwen2.5-3B [7], Qwen3-2B [6], LLaVA-OV [21], LFM-VL 2.5 [1], BioCLIP [42], SigLIP2 [46], and SCOLD [37]. Second, to establish standard image classification baselines, we fine-tuned representative vision-only architectures, comprising ResNet50 [14], EfficientNetV2S [44], DenseNet121 [16], and InceptionResNetV2 [43]. Lastly, we compared SOLAR directly against fine-tuned VLMs—namely CLIP [38], BioCLIP [42], SigLIP2 [46], AgriCLIP [29], SCOLD [37], and LLaVA 1.5 [21]—trained under identical conditions. All evaluations were conducted within a strictly consistent experimental framework to ensure fair comparison. Evaluation Metrics. We evaluate model performance using five complementary metrics. Accuracy (Acc) measures the proportion of correct predictions over the total number of test samples, reflecting the general reliability of the classification. F1-Score (F1) combines precision and recall into a single balanced measure to account for class-level performance. For generation quality, we employ ROUGELROUGE_L [24] complements these by measuring the longest common subsequence between generated and reference texts, thereby reflecting fluency and structural coherence. To evaluate the response quality of open-ended VQA output, we use GPTGPT that the LLM-as-a-judge [22] method with Gemini Flash 2.5 [9]. This evaluator compares the generated response with ground truth and rates it on a five-point scale based on semantic alignment, correctness, and completeness. Under this scheme, a score of 1 indicates a completely incorrect or irrelevant answer; 2 represents a partially relevant response with major inaccuracies; 3 denotes a generally correct output that lacks specific details; 4 signifies a fully correct and semantically equivalent answer; and 5 corresponds to an exceptionally precise, clear diagnosis that demonstrates a comprehensive understanding of both visual and textual contexts. The model is instructed to output only a single integer to facilitate structured data parsing. Together, these metrics provide a comprehensive assessment of both global correctness and per-class diagnostic effectiveness for classification tasks, as well as lexical fidelity and generation quality for generative task types. The detail prompt in the Table 2. Statistical Methods We compare between the best model and the second best by using a two-sided Wilcoxon signed-rank test. Performance scores are presented as the mean with 95%95\% confidence intervals (CIs) estimated via bootstrapping (n=1,000n=1,000 replicates). 4.3 Experimental Setup In this section, we describe the hardware and hyperparameter settings and evaluation method as below. used in our experiments. Hardware Platform: All experiments are performed on a local workstation with the following hardware and software specifications: CPU: Intel(R) Xeon(R) CPU E5-2686 v4 @ 2.30GHz, GPU: NVIDIA GeForce RTX 4060 Ti 16 GB, Memory: 128 GB RAM, System: Windows Hyperparameter Settings: For training, we run the model for 20 epochs using the AdamW optimizer with a learning rate of 1×10−41× 10^-4 and a weight decay of 0.01. The batch size is set to 32. For LoRA tuning, we apply the adapters to the query, key, value and output layers (Wq,Wk,Wv,WoW_q,W_k,W_v,W_o) of the language model with a rank r=16r=16, alpha α=32α=32, and a dropout rate of 0.1. For the MoE layer, the model has N=4N=4 experts and uses top-K=2K=2 routing. Each expert has a hidden size of 4096. The weights for the loss balance load and the loss of diversity are set to λ1=0.01 _1=0.01 and λ2=0.001 _2=0.001. Foundation Models: Both general-purpose and domain-specific vision-language foundation models are evaluated in a zero-shot multiple-choice setup. The prompt presents the query with four options. For single-label tasks (LC, HDC, PC, DC, SNC), each option corresponds to a distinct category. For the multi-label Symptom Identification (SI) task, since a single image may have multiple distinct symptoms, we ensure that each option contains only a single symptom associated with that image. Vision Models: We add a classification layer to each model for each question type. The number of classification heads corresponds to the number of classes for each type of question. Fine-tuned VLMs: Contrastive vision-language models are fine-tuned on the training set. During inference, option text prompts corresponding to all class names within a task are encoded through the text encoder. For a given input image, the cosine similarity between the image embedding and all option text embeddings is computed. The option that yields the maximal similarity score is selected as the predicted label. Unlike traditional image classification models, SOLAR directly generates natural language answers. For single-label tasks (LC, HDC, PC, DC, SNC), predicted labels are extracted from the generated text via exact keyword matching against the ground truth. For the multi-label Symptom Identification (SI) task, predicted symptom terms are parsed into a discrete set and evaluated against the ground-truth symptom set. 5 Experimental Result 5.1 Closed Visual Question Answering Performance To evaluate the effectiveness of SOLAR, we conduct comprehensive comparative experiments against three categories of models on our internal dataset and three external datasets. SOLAR demonstrated superior performance on all VQA tasks. On average, SOLAR consistently outperformed all SOTA models (Figure 3a). In particular, SOLAR significantly outperformed all competing models in basic perception tasks such as HDC (93.7%93.7\%, 95%95\% CI: 91.3%91.3\%–96.0%96.0\%), pathology understanding tasks such as PC (64.0%64.0\%, 95%95\% CI: 61.5%61.5\%–66.5%66.5\%) and SI (53.6%53.6\%, 95%95\% CI: 50.0%50.0\%–57.0%57.0\%), and disease diagnosis tasks including DC (57.4%57.4\%, 95%95\% CI: 54.5%54.5\%–60.3%60.3\%) and SNC (61.8%61.8\%, 95%95\% CI: 59.0%59.0\%–64.8%64.8\%). SOLAR achieved an average increase of +2.15+2.15% (P≤0.0001P≤ 0.0001), with respective performance differences of +3.49+3.49% (HDC), +0.72+0.72% (PC), +4.63+4.63% (SI), +4.58+4.58% (DC), +2.37+2.37% (SNC), and −4.36-4.36% (LC) compared with the second-best model (P≤0.0001P≤ 0.0001). In addition, we compare our model with the top-3 models from each model category across all six tasks (Figure 3b-d, and Tables 6-9). Specifically, when comparing with the top-3 fine-tuned VLMs, SOLAR outperforms on almost all tasks with +16.79%+16.79\% and +0.196+0.196 (LLaVA), +37.68%+37.68\% and +0.263+0.263 (SigLIP2), and +60.70%+60.70\% and +0.513+0.513 (BioCLIP), but underperforms on LC with a gap of −3.95%-3.95\% and −0.011-0.011 (LLaVA). Similarly, SOLAR also outperforms vision-only models in almost all tasks with +52.37%+52.37\% and +0.208+0.208 (EfficientNetV2S), +69.34%+69.34\% and +0.496+0.496 (ResNet50), and +70.13%+70.13\% and +0.503+0.503 (DenseNet121), but slightly lower on LC with gaps of −2.39%-2.39\% (EfficientNetV2S), −5.81%-5.81\% (ResNet50), and −2.28%-2.28\% (DenseNet121). Specifically, with top-1 representatives from each category, SOLAR also exceeded other models in almost all tasks, but still underperformed on LC with −3.95%-3.95\% and −0.011-0.011 (LLaVA) and −4.34%-4.34\% and −0.027-0.027 (Qwen2.5 3B). Figure 3: Performance comparison across the six VQA tasks and the overall average and performance comparison of the top-performing models in each category across the six tasks. a, The box plots show the distribution of accuracy for SOLAR with the best models in each category(LLaVA, EfficientNetV2S, and Qwen2.5 3B) across the six tasks (HDC, DC, SNC, PC, SI and LC). b, Comparison of SOLAR with the fine-tuned VLMs achieving the highest average performance(LLaVA, SigLIP2, BioCLIP). c, Comparison of SOLAR with the CNN architectures achieving the highest average performance(EfficientNetV2S, ResNet50, DenseNet121). d, Comparison of the SOLAR model with the best-performing models in each evaluation domain (LLaVA, EfficientNetV2S, Qwen2.5 3B). Error bar represented as the mean with 95% CIs estimated using the bootstrap method (n=1,000n=1,000 replicates). The two-sided Wilcoxon signed-rank test was used to assess the statistical differences between SOLAR and the second-best model. P∗∗≤0.01^**P≤ 0.01, ∗P≤0.001^***P≤ 0.001, ∗P≤0.0001^****P≤ 0.0001. (purple indicates Acc, and pink denotes F1) SOLAR significantly outperformed baselines on internal and external datasets. Figure 4 demonstrated that SOLAR consistently outperformed all competing models, leading with the highest average accuracy of 69.08%69.08\% (95%95\% CI: 67.88%67.88\%–70.21%70.21\%). This represents an accuracy improvement of +8.32%+8.32\% over EfficientNetV2S, +10.58%+10.58\% over ResNet50, and +10.59%+10.59\% over SigLIP2 (Table 1). Specifically, on our internal dataset and TLD-3 dataset, SOLAR maintained superior performance, achieving accuracies of 97.68%97.68\% (95%95\% CI: 96.50%96.50\%–98.83%98.83\%) and 91.67%91.67\% (95%95\% CI: 89.33%89.33\%–93.67%93.67\%), respectively. In these specific evaluations, SOLAR outperformed LLaVA by +1.78%+1.78\%, EfficientNet by +10.24%+10.24\%, and Qwen2.5 3B by +10.45%+10.45\% (Figure 4a-d). Specifically, when comparing with top-3 fine-tuned VLMs(Figure 4e), SOLAR covers almost all other VLMs +8.55%+8.55\% and +0.123+0.123 (LLaVA), +42.35%+42.35\% and +0.348+0.348 (SigLIP2), and +44.72%+44.72\% and +0.385+0.385 (BioCLIP). Meanwhile, SOLAR still outperformed vision models in almost dataset (Figure 4f) such as +37.38%+37.38\% and +0.222+0.222 (EfficientNetV2S), +40.91%+40.91\% and +0.349+0.349 (ResNet50), and +43.52%+43.52\% and +0.334+0.334 (DenseNet121), but underperformed in TLID, with the gaps are −4.08%-4.08\% and −0.093-0.093 (EfficientNetV2S). (See detail in Tables 6-9). Figure 4: Average accuracy comparison of SOLAR and baseline models across the four evaluation datasets including (a)TomaMMU, (b) TLID (c) TLD-3 and (d) TomaAD datasets. The bars represent the mean accuracy of SOLAR with the best models in each category (LLaVA, EfficientNetV2S, and Qwen2.5 3B). Performance comparison on tomato disease VQA datasets. e, Comparison of SOLAR with the fine-tuned VLMs achieving the highest average performance(LLaVA, SigLIP2, BioCLIP) f, Comparison of SOLAR with the CNN architectures achieving the highest average performance (EfficientNetV2S, ResNet50, DenseNet121). Error bar represented as the mean with 95% CIs estimated using the bootstrap method (n=1,000n=1,000 replicates). The two-sided Wilcoxon signed-rank test was used to assess the statistical differences between SOLAR and the second-best model. P∗∗≤0.01^**P≤ 0.01, ∗P≤0.001^***P≤ 0.001, ∗P≤0.0001^****P≤ 0.0001. (purple indicates Acc, and pink denotes F1) SOLAR demonstrates strong generative capabilities for agricultural VQA. Because agricultural VQA requires not only accurate classification but also informative context, we further evaluated the models’ capacities to generate free-form responses. As shown on Table 1, on average, it outperformed Qwen3-2B by +0.024+0.024 (ROUGELROUGE_L) and +0.759+0.759 (GPT), with consistent improvements over InternVL3-2B (+0.028+0.028, +0.674+0.674) and InternVL3-1B (+0.034+0.034, +1.013+1.013). Specifically, on our internal dataset, SOLAR achieved the highest performance (ROUGELROUGE_L: 0.9950.995; GPT: 4.7414.741), outperforming the strongest competitors, including LFM-VL 2.5 (by +0.06+0.06 and +1.000+1.000, respectively), Qwen3-2B (+0.063+0.063, +1.312+1.312), and InternVL3-2B (+0.024+0.024, +1.371+1.371). Conversely, on the TLID dataset, SOLAR trailed the two leading models, yielding a ROUGELROUGE_L of 0.8940.894 and a GPT score of 2.6492.649. This represented a deficit of 0.0360.036 (ROUGELROUGE_L) and 0.4620.462 (GPT) compared to Qwen3-2B, alongside gaps of 0.0250.025 and 0.670.67 relative to InternVL3-2B. However, SOLAR regained its lead on the TLD-3 and TomaAD datasets, scoring 0.9730.973 and 0.9200.920 in ROUGELROUGE_L, and 4.5494.549 and 3.3343.334 in GPT, respectively. Across these datasets, SOLAR maintained substantial margins over Qwen3-2B (by +0.068+0.068 ROUGELROUGE_L and +2.186+2.186 GPT), InternVL3-2B (+0.072+0.072 and 1.9971.997), and InternVL3-1B (+0.079+0.079 and +2.431+2.431). Overall, SOLAR achieved the highest average scores across all evaluated datasets (ROUGELROUGE_L: 0.9460.946; GPT: 3.8183.818). (See detail in Tables 10-13) Table 1: Generative evaluation of SOLAR and baseline models across four tomato disease VQA datasets, measured by ROUGELROUGE_L and GPTGPT judged by Gemini Flash 2.5. Model TomaMMU TLID TLD-3 TomaAD Average ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT InternVL3 1B [55] 0.926 3.311 0.908 2.459 0.926 3.311 0.888 2.141 0.912 2.805 InternVL3 2B [55] 0.931 3.370 0.919 3.319 0.931 3.370 0.890 2.516 0.918 3.144 Qwen2.5 3B [7] 0.931 3.382 0.909 2.750 0.931 3.382 0.882 2.252 0.913 2.941 Qwen3 2B [6] 0.932 3.429 0.930 3.111 0.932 3.429 0.893 2.268 0.922 3.059 LlaVA-OV [21] 0.921 3.153 0.906 2.522 0.921 3.153 0.885 2.347 0.908 2.794 LFM-VL 2.5 [1] 0.935 3.471 0.884 2.272 0.935 3.471 0.862 1.835 0.904 2.762 SOLAR 0.995 4.741 0.894 2.649 0.973 4.549 0.920 3.334 0.946 3.818 5.2 Interpretability Learned representations from SOLAR. To qualitatively evaluate the representation learning capability of SOLAR, we analyzed the UMAP visualizations of its learned visual feature embeddings (Figure 5). Figure 5a establishes original image embeddings, which then resolve into highly distinct, task-specific manifolds when categorized across the six VQA tasks by the MoE layer (Figure 5b). In particular, basic perceptual tasks such as HDC and LC formed well-separated binary decision boundaries by SOLAR, demonstrating robust linear separability (Figure 5c-d). When processing biological features, the PC task yielded compact clusters corresponding to distinct pathogen types (Figure 5e). For high-precision diagnostic outputs, both DC and SNC exhibited tight intra-class compactness and clear inter-class margins across 15 tomato disease categories (Figure 5f) and 12 scientific nomenclatures (Figure 5g). This indicates that SOLAR effectively grounds expert pathology logic into distinct feature spaces. Lastly, on the most challenging Symptom Identification (SI) task, SOLAR successfully partitioned 26 fine-grained symptoms (Figure 5h), evidencing precise, localized feature extraction despite high semantic overlap among visual symptom presentations. Figure 5: UMAP visualizations of learned visual features. a, UMAP of image embeddings before the MoE layer. b, UMAP by the 6 VQA tasks. c, UMAP of the HDC task. d, UMAP of the LC task. e, UMAP of the PC task. f, UMAP of the DC task. g, UMAP of the SNC task. h, UMAP of the SI task. Expert Routing for multiple VQA tasks. We next hypothesized that configuring an MoE architecture with n=6n=6 experts and a Top-1 hard routing mechanism would optimally map the six distinct question types in our dataset (HDC, DC, LC, SI, PC, and SNC) to the experts, thereby allowing each expert to specialize exclusively in a single task. However, this configuration yielded the lowest overall performance, achieving an average accuracy of 43.46%43.46\% and an F1 of 0.4680.468 (Table 3). This severe performance degradation stems from the semantic interdependence inherent to domain-specific multi-task VQA. For instance, answering a scientific nomenclature (SNC) question intrinsically relies on recognizing both the visual disease category (DC) and related localized symptoms (SI). Enforcing a fixed routing path completely isolated the specialized experts, precluding essential cross-task knowledge transfer. Consequently, each expert was forced to independently learn visual features and textual reasoning from an artificially narrowed data partition. This architectural bottleneck prevented the model from capturing generalized foundational representations and ultimately precipitated severe overfitting. Figure 6: Expert routing distribution heatmaps for image-level MoE with different configurations of expert numbers: n=4n=4 (a), n=6n=6 (b), and n=8n=8 (c). Each cell indicates the average gating probability of a specific question type routed to the corresponding expert Qualitative analysis and hierarchical error propagation. To interpret the diagnostic reasoning of SOLAR, we analyzed the hierarchical flow of its predictions across the sequential VQA tasks (Figure 7a). The alluvial diagram illustrates the propagation of diagnostic logic, revealing that SOLAR maintains exceptional accuracy on foundational perceptual and biological tasks, such as Leaf Counting (99.1%99.1\%), Healthy/Disease Classification (98.8%98.8\%), and Pathogen Classification (99.5%99.5\%). However, Symptom Identification (SI) emerges as the primary bottleneck, containing the highest proportion of misclassifications (7.8%7.8\%). Because the agricultural VQA tasks are semantically interdependent, errors at the intermediate symptom level naturally cascade into downstream failures in Disease Classification (DC) and Scientific Nomenclature Classification (SNC). This hierarchical dependence is further exemplified through qualitative comparison (Figure 7b). When processing a representative pathological sample, all evaluated models successfully resolved the basic perceptual tasks (LC: “One”; HDC: “Disease”). However, the baseline architectures (LLaVA, EfficientNetV2S, and Qwen2.5-3B) failed to extract the correct localized features, leading to incorrect symptom and pathogen identification and subsequent misdiagnoses (e.g., incorrectly predicting Late Blight or Bacterial Blight). In contrast, SOLAR explicitly identified the correct fine-grained symptoms (“concentric ring, yellowing”) and pathogen type (“Fungal”). By correctly anchoring its reasoning in these intermediate biological features, SOLAR successfully deduced the specific disease (“Early Blight”) and its corresponding scientific nomenclature (Alternaria linariae). This demonstrates a transparent, step-by-step diagnostic progression that mirrors the hierarchical logic of an expert pathologist. Figure 7: Diagnostic error propagation and qualitative performance of SOLAR across VQA tasks. a, Alluvial diagram tracking prediction accuracy and error propagation flows across the six sequential VQA tasks (LC → SNC) for on TomaMMU. Teal bands represent correct predictions, while red bands indicate incorrect predictions. The flow thicknesses correspond to the proportion of samples, highlighting SOLAR’s near-perfect accuracy on Pathogen Classification (PC, 99.5%99.5\% correct) and demonstrating that Symptom Identification (SI) remains the most challenging bottleneck, harboring the highest error rate (7.8%7.8\%). b, Qualitative comparative analysis of model predictions for a representative pathological leaf sample. Responses from SOLAR (purple) are contrasted against LLaVA (light blue), EfficientNetV2S (gray-blue), and Qwen2.5-3B (orange) across all six tasks. While all models correctly identify the perceptual Leaf Count (LC: "One") and basic Healthy/Disease Classification (HDC: "Disease"), only SOLAR successfully correlates the fine-grained visual symptoms ("concentric ring, yellowing") to the correct Disease Classification (DC: "Early Blight") and corresponding Scientific Nomenclature Classification (SNC: Alternaria linariae). Conversely, the baseline models exhibit cascading errors, misclassifying the sample as Late Blight (Phytophthora infestans) or Bacterial Blight. 5.3 Ablation Study To isolate the contributions of key architectural components within SOLAR, we conducted comprehensive ablation studies across five dimensions: (1) the individual impact of core modules, (2) the effectiveness of different vision encoders, (3) the scaling of the large language model (LLM) layers, (4) the positioning of the MoE layer, and (5) the optimal number of experts. To ensure a rigorous and fair comparison, all experiments were executed under identical configurations and evaluation protocols, as detailed in Section 4. Figure 8: Ablation studies and analysis of the SOLAR architecture. a, Performance of SOLAR when removing individual key components. b, Performance of SOLAR when employing different vision encoder backbones (CLIP, BioCLIP, AgriCLIP, and SCOLD). c, Performance of SOLAR when employing different LLM decoders (BioGPT, Qwen3, and LFM-VL 2.5). Effectiveness of core modules in SOLAR. To delineate the individual contributions of SOLAR’s architectural components to overall diagnostic performance, we evaluated three configurations: the isolated removal of the Mixture of Experts layer (w/o MoE), the removal of the fusion adapter (w/o Fusion Adapter), and the complete SOLAR architecture (Figure 8a and detail numbers in Table 3 for each dataset). Specifically, across all four datasets, the fully integrated SOLAR model secured the highest average performance (69.06%69.06\% Acc, 0.6660.666 F1), exceeding the w/o Fusion Adapter by +0.33%+0.33\% and +0.002+0.002, and the w/o MoE by +2.86%+2.86\% and +0.024+0.024. This confirms that the synergistic integration of both the MoE and fusion layers is essential for maximizing generalized diagnostic capabilities. (See detail in Tables 26-29) Performance of Different Vision Encoders. To assess the impact of the visual feature extractor on diagnostic capability, we substituted SOLAR’s default SCOLD encoder with three widely used vision backbones: CLIP, BioCLIP, and AgriCLIP (Figure 8b and Table 4). As demonstrated in Figure 8b, substituting SCOLD with less specialized encoders resulted in a marked decline in overall diagnostic performance. The specialized plant disease encoder (SCOLD) achieved the highest overall results, reaching 69.06%69.06\% Acc and a 66.60%66.60\% F1. In comparison, the general-domain model (CLIP) achieved 65.61%65.61\% Acc and 63.90%63.90\% F1, while the macro-agricultural model (AgriCLIP) scored 65.04%65.04\% Acc and 63.40%63.40\% F1. The taxonomy-focused biology model (BioCLIP) exhibited the lowest performance (62.76%62.76\% Acc, 62.80%62.80\% F1). These results indicate that general or broadly agricultural models lack the fine-grained perceptual capabilities required for precise symptom identification. SCOLD’s domain-specific pre-training effectively captures the key details of foliar diseases, yielding superior foundational representations for downstream pathology tasks. (See detail in Tables 14-17) Performance of different LLMs. To elucidate the impact of the underlying large language model (LLM) backbone on diagnostic reasoning, we compared SOLAR’s default textual decoder, LFM-VL 2.5, against two SOTA generative models: Qwen3 [6] and BioGPT[26] (Figure 8c and Table 5). As illustrated in Figure 8c, substituting the text decoder significantly impacts diagnostic efficacy across different parameter scales. SOLAR with LFM-VL 2.5 (1,171M parameters) achieved the highest overall performance, reaching 69.06%69.06\% Acc and a 66.60%66.60\% F1. This configuration yielded substantial performance gains of +1.55%+1.55\% in Acc and +1.50%+1.50\% in F1 compared to the BioGPT backbone (350M parameters), alongside respective improvements of +1.99%+1.99\% and +0.80%+0.80\% over the Qwen3 backbone (601M parameters). This demonstrates that LFM-VL 2.5’s targeted multimodal alignment and architectural efficiency provide superior diagnostic reasoning capabilities compared to purely general-domain or tangentially related medical models. (See detail in Tables 18-21) MoE layer positioning. To determine the optimal integration flow for expert selection, we evaluated three placements for the Mixture of Experts (MoE) module (Table 2). Dual MoE employs independent MoE blocks on the image and text branches prior to combination. Fusion MoE applies a single MoE module late in the network, processing the already-mixed multimodal features. Finally, our proposed Image MoE places the module exclusively on the visual branch, uniquely utilizing the textual question as the gating signal to route visual features. While the Dual MoE configuration marginally outperformed Image MoE on the TomaMMU (by +0.57%+0.57\% Acc, +0.005+0.005 F1) and TLD-3 (by +0.34%+0.34\% Acc, +0.004+0.004 F1) datasets, Image MoE proved substantially more robust across diverse and challenging diagnostic scenarios. On the TLID dataset, Image MoE (33.70%33.70\% Acc, 0.3350.335 F1) exceeded Dual MoE by +0.87%+0.87\% (Acc) and +0.008+0.008 (F1), and Fusion MoE by +0.90%+0.90\% and +0.009+0.009. This advantage widened significantly on the TomaAD dataset (53.20%53.20\% Acc, 0.5010.501 F1), where Image MoE outperformed Fusion MoE by +3.81%+3.81\% and +0.030+0.030, and Dual MoE by a substantial +7.85%+7.85\% and +0.061+0.061. Overall, the question-gated Image MoE strategy achieved the highest average performance (69.06%69.06\% Acc, 0.6660.666 F1), yielding net improvements of +1.36%+1.36\% and +0.011+0.011 over Fusion MoE, and +1.95%+1.95\% and +0.015+0.015 over Dual MoE. This indicates that using linguistic cues to dynamically route foundational visual features is more effective than fully independent or post-fusion routing. (See detail in Tables 22-25) Table 2: Performance of different MoE placement strategies across four datasets in the SOLAR architecture TomaMMU TLID TLD-3 TomaAD Average Type Acc (%) F1 Acc (%) F1 Acc (%) F1 Acc (%) F1 Acc (%) F1 Dual MoE 98.25 0.979 32.83 0.327 92.01 0.859 45.35 0.440 67.11 0.651 Fusion MoE 97.02 0.966 32.80 0.326 91.60 0.855 49.39 0.471 67.70 0.655 Image MoE 97.68 0.974 33.70 0.335 91.67 0.855 53.20 0.501 69.06 0.666 Number of experts. To determine the optimal capacity of the Image MoE layer, we evaluated the impact of varying the total number of experts (n∈4,6,8n∈\4,6,8\) while maintaining a constant Top-22 routing mechanism (k=2k=2) (Table 3). Top-22 routing restricts the gating network to select and aggregate outputs from only the two highest-scoring experts per input image. This ensures a stable, bounded computational cost while promoting collaborative feature sharing between the activated experts. Focusing on these Top-22 configurations, our default setting of n=4n=4 experts achieved the highest overall performance, with 69.06%69.06\% Acc and 0.6660.666 F1 (+1.56%+1.56\% Acc and −0.031-0.031 F1 relative to n=6n=6). Specifically, the n=4n=4 configuration maximized performance on the TomaMMU (97.68%97.68\% Acc; +0.28%+0.28\% over n=6n=6) and TomaAD (53.20%53.20\% Acc; +6.69%+6.69\% over n=6n=6) datasets. Conversely, on the TLID and TLD-3 datasets, n=4n=4 yielded slightly lower accuracies of 33.70%33.70\% and 91.67%91.67\%, trailing the n=6n=6 configuration by margins of −0.53%-0.53\% and −0.37%-0.37\%, respectively. The general performance degradation observed with larger expert pools (n≥6n≥ 6) suggests that excessive capacity exacerbates overfitting on limited training data and increases the gating network’s optimization complexity. As visualized in Figure 6, a moderate number of experts (n=4n=4) encourages distinct functional specialization. In contrast, an over-parameterized MoE layer results in dispersed routing patterns, load imbalances, and a failure of individual experts to develop localized diagnostic specializations. (See detail in Tables 30-33) Table 3: Ablation study on the number of experts in the Image MoE layer of the SOLAR architecture TomaMMU TLID TLD-3 TomaAD Average Expert Top-k Acc (%) F1 Acc (%) F1 Acc (%) F1 Acc (%) F1 Acc (%) F1 4 2 97.68 0.974 33.70 0.335 91.67 0.855 53.20 0.501 69.06 0.666 6 1 96.81 0.965 7.85 0.079 52.63 0.493 33.65 0.337 47.73 0.468 6 2 97.23 0.968 34.23 0.369 92.04 0.902 46.51 0.548 67.50 0.697 8 2 97.40 0.972 33.03 0.330 91.93 0.860 45.24 0.440 66.90 0.650 6 Discussion In this study, we present SOLAR, a generative multimodal foundation model for a comprehensive understanding of tomato diseases. Through extensive benchmark evaluation across six hierarchical VQA tasks, we show that SOLAR achieves superior performance over existing vision and vision-language models for basic perception, pathology understanding, and precision disease diagnosis. Importantly, in contrast to previous studies that relied on isolated, single-label image classification, we leveraged a generative framework to capture the complementary semantic relationships between visual symptoms and diagnostic reasoning. This multimodal approach successfully mirrors the cognitive, coarse-to-fine diagnostic workflow of an agricultural expert. The performance gain achieved by SOLAR is largely attributable to its Expert Fusion module, which dynamically routes and integrates visual features based on the semantic context of the input question. Existing agricultural studies have predominantly used static, off-the-shelf foundation models or simple projectors that impose a rigid mapping between image and text spaces [21, 47]. By contrast, SOLAR is custom-designed with a question-conditioned Image MoE layer. This module allows the network to adaptively emphasize diagnostically relevant visual regions—such as distinct lesion patterns or necrosis—depending on the specific nature of the query. This mechanism effectively prevents the overlap of learned features and mitigates severe overfitting on limited domain data. The semantic interdependence among agricultural tasks poses a major challenge for training reliable deep learning models. For instance, answering a highly specific SNC question that uses scientific nomenclature inherently relies on accurately recognizing both the visual disease category and localized leaf symptoms. Our approach provides an effective solution to this problem by using a Top-2 MoE routing mechanism across four experts, which maintains stable computational cost while encouraging distinct functional specialization among the experts. This training paradigm can be extended and applied in building interpretable multimodal foundation models for other complex agricultural domains. Accurate and explainable disease diagnosis has significant practical implications for precision agriculture. There are important conceptual and practical distinctions between basic binary disease detection and the comprehensive understanding of pathology that is the primary focus of most existing agricultural classifiers. Given that simple classifiers only provide a black-box label without contextual grounding, their impact in real-world farming is limited to a basic assistive role. However, fine-grained symptom identification is a much more challenging problem due to the high semantic overlap and substantial intra-class visual variability in disease manifestations. By explicitly anchoring its reasoning in intermediate biological features (such as concentric rings or yellowing), the multimodal SOLAR model significantly improved upon traditional classification pipelines, demonstrating a transparent, step-by-step diagnostic progression. Despite its strong overall performance, symptom identification task remains the most difficult bottleneck in the diagnostic pipeline, harboring the highest error rate (7.8%7.8\%) in our hierarchical analysis. Furthermore, SOLAR exhibited vulnerabilities when generalizing to highly complex, in-the-wild datasets such as TLID. While the model excels in carefully curated environments like TomaMMU and TomaAD, it underperformed compared to large general-domain models (e.g., Qwen3-2B) on the generative tasks in the TLID dataset. This discrepancy highlights the inherent trade-off of using a hyper-specialized plant disease encoder such as SCOLD. In real-world greenhouse environments characterized by chaotic backgrounds, multiple co-infections, and nutrient deficiencies, highly specialized encoders can become overly sensitive to localized pathological artifacts, whereas massive general models maintain a more robust macroscopic understanding of overall plant health. Although the results of SOLAR’s generative VQA framework are highly promising, they were evaluated primarily on isolated leaf images from specific datasets. Before the model can be considered for widespread implementation in commercial greenhouses, several steps are needed to ensure rigorous evaluation of its robustness and field utility. First, these findings should be validated in future studies with larger, multi-regional datasets that capture extreme environmental variations and complex multi-disease interactions. Second, expanding multimodal inputs to integrate external metadata, such as climate conditions and temporal growth stages, will be required to support precision agricultural management fully. Lastly, while SOLAR excels at identifying diseases and symptoms, translating these diagnostic insights into actionable, localized treatment protocols remains an open challenge. Future iterations must bridge the gap between diagnostic and prescriptive agriculture by generating safe, evidence-based intervention strategies tailored to the farmer’s specific environmental constraints. 7 Conclusion In this study, we introduced SOLAR, a generative multimodal architecture that fundamentally reformulates tomato leaf disease diagnosis as a hierarchical Visual Question Answering task. By integrating a question-conditioned fusion Mixture-of-Experts layer with a residual expert fusion mechanism, SOLAR dynamically routes and aligns fine-grained visual features based on the specific semantic context of the diagnostic query. Extensive evaluations across four diverse datasets demonstrate that SOLAR consistently achieves state-of-the-art performance, surpassing existing vision-only architectures, general vision-language models, and domain-specific foundation models in both classification accuracy and generative response quality. Crucially, by mirroring the step-by-step diagnostic reasoning of agricultural experts, SOLAR overcomes the interpretability limitations of traditional "black-box" classifiers, providing a transparent and contextually grounded understanding of pathology. These findings validate the efficacy of dynamic multimodal soft-routing for complex agricultural VQA tasks. Ultimately, SOLAR establishes a highly scalable and explainable framework for precision agriculture, paving the way for next-generation, expert-level disease management systems deployed in real-world farming environments. Acknowledgment This research was supported by the Hyundai Motor Chung Mong-Koo Foundation Global Scholarship (GSS-25-02120). Data Availability TomaMMU, TomaBench, TLID, TomaAD, and TLD-3 are publicly available and can be accessed from the following: TomaMMU and TomaBench (https://huggingface.co/datasets/enalis/TomaMMU), TLID (https://data.mendeley.com/datasets/kt64b2kh89/2), TLD-3 (https://data.mendeley.com/datasets/zfv4j7855/1) and TomaAD (https://doi.org/10.6084/m9.figshare.31145998). CRediT authorship contribution statement Khang Nguyen Quoc contributed to conceptualization, data curation, methodology, software, model development, data analysis and visualization, and writing – original draft. Minh-Phuoc Tran contributed to methodology, data curation, software, model development, data analysis, visualization, and writing – original draft. Gia-Han Truong contributed to data curation, validation, and writing – review & editing. Luyl-Da Quach contributed to methodology, project administration, supervision, validation, and writing – review & editing. Declaration of Generative AI Use The authors acknowledge the use of Claude Sonnet 5 and Grammarly to enhance the grammar and clarity of the manuscript. After using them, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. Declaration of Competing Interest The authors declare no known competing conflicts of interest that could have influenced the work reported in this paper. References [1] L. AI (2025) LFM2 technical report. arXiv preprint arXiv:2511.23404. Cited by: §3.2, §4.2, Table 1. [2] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. L. Zitnick, and D. Parikh (2015) VQA: visual question answering. In 2015 IEEE International Conference on Computer Vision (ICCV), p. 2425–2433. External Links: Link, Document Cited by: §2.3. [3] M. A. Arshad, T. Z. Jubery, T. Roy, R. Nassiri, A. K. Singh, A. Singh, C. Hegde, B. Ganapathysubramanian, A. Balu, A. Krishnamurthy, and S. Sarkar (2024) Leveraging vision language models for specialized agricultural tasks. arXiv. External Links: Document, Link Cited by: §2.2. [4] M. Astani, M. Hasheminejad, and M. Vaghefi (2022) A diverse ensemble classifier for tomato disease recognition. Computers and Electronics in Agriculture 198, p. 107054. External Links: Document Cited by: §1. [5] M. Awais, A. H. Salem Abdulla Alharthi, A. Kumar, H. Cholakkal, and R. M. Anwer (2025) AgroGPT : efficient agricultural vision-language model with expert tuning. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), p. 5687–5696. External Links: Link, Document Cited by: §2.2. [6] S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, et al. (2025) Qwen3-vl technical report. arXiv preprint arXiv:2511.21631. Cited by: §4.2, §5.3, Table 1. [7] S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y. Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y. Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin (2025) Qwen2.5-vl technical report. External Links: 2502.13923, Link Cited by: §4.2, Table 1. [8] Y. A. Bezabh, A. M. Ayalew, B. M. Abuhayi, T. N. Demlie, E. A. Awoke, and T. E. Mengistu (2024) Classification of mango disease using ensemble convolutional neural network. Smart Agricultural Technology 8, p. 100476. External Links: ISSN 2772-3755, Link, Document Cited by: §2.1. [9] G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, and orthers (2025) Gemini 2.5: pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. External Links: 2507.06261, Link Cited by: §4.2. [10] Directorate-General for Agriculture and Rural Development (2026)Tomato dashboard – information on the agri-food data portal(Website) European Commission. External Links: Link Cited by: §1. [11] W. Fedus, B. Zoph, and N. Shazeer (2021) Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity. arXiv e-prints, p. arXiv:2101.03961. External Links: Document, 2101.03961 Cited by: §3.4. [12] W. Fedus, B. Zoph, and N. Shazeer (2022) Switch transformers: scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research 23 (120), p. 1–39. Cited by: §3.2. [13] A. Gauba, I. Pi, Y. Man, Z. Pang, V. Adve, and Y. Wang (2026) Agmmu: a comprehensive agricultural multimodal understanding benchmark. Advances in Neural Information Processing Systems 38. Cited by: §2.3. [14] K. He, X. Zhang, S. Ren, and J. Sun (2016) Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 770–778. Cited by: §4.2. [15] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen (2021) LoRA: low-rank adaptation of large language models. External Links: 2106.09685, Link Cited by: §3.4. [16] F. Iandola, M. Moskewicz, S. Karayev, R. Girshick, T. Darrell, and K. Keutzer (2014) Densenet: implementing efficient convnet descriptor pyramids. arXiv preprint arXiv:1404.1869. Cited by: §4.2. [17] A. Jafar, N. Bibi, R. A. Naqvi, A. Sadeghi-Niaraki, and D. Jeong (2024) Revolutionizing agriculture with artificial intelligence: plant disease detection methods, applications, and their limitations. Frontiers in Plant Science 15, p. 1356260. External Links: Document Cited by: §1. [18] D. S. Joseph, P. M. Pawar, and K. Chakradeo (2024) Real-time plant disease dataset development and detection of plant disease using deep learning. IEEE Access 12, p. 16310–16333. External Links: ISSN 2169-3536, Link, Document Cited by: §2.1. [19] H. Kim (2026) Geometric regularization in mixture-of-experts: the disconnect between weights and activations. External Links: 2601.00457, Link Cited by: §3.4. [20] Y. Lan, Y. Guo, Q. Chen, S. Lin, Y. Chen, and X. Deng (2023) Visual question answering model for fruit tree disease decision-making based on multimodal deep learning. Frontiers in Plant Science 13. External Links: ISSN 1664-462X, Link, Document Cited by: §2.3. [21] B. Li, Y. Zhang, D. Guo, R. Zhang, F. Li, H. Zhang, K. Zhang, P. Zhang, Y. Li, Z. Liu, et al. (2024) Llava-onevision: easy visual task transfer. arXiv preprint arXiv:2408.03326. Cited by: §3.2, §4.2, Table 1, §6. [22] H. Li, Q. Dong, J. Chen, H. Su, Y. Zhou, Q. Ai, Z. Ye, and Y. Liu (2024) LLMs-as-judges: a comprehensive survey on llm-based evaluation methods. External Links: 2412.05579, Link Cited by: §4.2. [23] H. Li, H. Wu, Q. Li, and C. Zhao (2025) A review on enhancing agricultural intelligence with large language models. Artificial Intelligence in Agriculture 15 (4), p. 671–685. External Links: ISSN 2589-7217, Link, Document Cited by: §2.2. [24] C. Lin (2004) Rouge: a package for automatic evaluation of summaries. In Text summarization branches out, p. 74–81. Cited by: §4.2. [25] J. Lin, G. Hu, and J. Chen (2025) Mixed data augmentation and osprey search strategy for enhancing yolo in tomato disease, pest, and weed detection. Expert Systems with Applications 264, p. 125737. External Links: Document Cited by: §1, §2.1. [26] R. Luo, L. Sun, Y. Xia, T. Qin, S. Zhang, H. Poon, and T. Liu (2022) BioGPT: generative pre-trained transformer for biomedical text generation and mining. Briefings in Bioinformatics 23 (6). External Links: ISSN 1477-4054, Link, Document Cited by: §5.3. [27] A. Nanavaty, R. Sharma, B. Pandita, O. Goyal, S. Rallapalli, M. Mandal, V. K. Singh, P. Narang, and V. Chamola (2024) Integrating deep learning for visual question answering in agricultural disease diagnostics: case study of wheat rust. Scientific Reports 14 (1). External Links: ISSN 2045-2322, Link, Document Cited by: §2.3. [28] S. A. Nawaz, J. Li, U. A. Bhatti, M. U. Shoukat, and R. M. Ahmad (2022) AI-based object detection latest trends in remote sensing, multimedia and agriculture applications. Frontiers in Plant Science 13, p. 1041514. External Links: Document Cited by: §1. [29] U. Nawaz, A. Muhammad, H. Gani, M. Naseer, F. S. Khan, S. Khan, and R. Anwer (2025) AgriCLIP: adapting clip for agriculture and livestock via domain-specialized cross-model alignment. In Proceedings of the 31st International Conference on Computational Linguistics, p. 9630–9639. Cited by: §4.2. [30] K. H. Nguyen, H. V. N. Nguyen, H. N. Tran, and L. Quach (2023) Combining autoencoder and yolov6 model for classification and disease detection in chickens. In Proceedings of the 2023 8th International Conference on Intelligent Information Technology, ICIIT 2023, p. 132–138. External Links: Link, Document Cited by: §1. [31] K. Q. Nguyen, H. C. G. Nguyen, T. N. D. Le, V. B. Q. Thai, B. P. Tran, L. T. T. Le, and L. Quach (2024) Tiny-cnn: structuring convolutional neural networks for accurate classification of rice leaf diseases in resource-constrained environments. In Proceedings of the 2024 9th International Conference on Intelligent Information Technology, ICIIT ’24, New York, NY, USA, p. 171–178. External Links: ISBN 9798400716713, Link, Document Cited by: §1. [32] T. V. Nguyen, K. Nguyen Quoc, D. Harwath, L. Quach, and P. D. Dao (2026) A multiregional image–text dataset and benchmark for vision-language modeling of plant diseases. bioRxiv, p. 2026–07. Cited by: §2.3. [33] O. Prabhu, M. T, and S. Shetty (2025) An investigation on advances in transfer learning and explainable ai for mango leaf disease detection. Smart Agricultural Technology 12, p. 101348. External Links: ISSN 2772-3755, Link, Document Cited by: §2.1. [34] L. Quach, K. N. Quoc, C. Nguyen, and N. Thai-Nghe (2024) Explainable ai for plant disease detection: assessing explainability in classifying maize leaves diseases with focus score and ablation-cam. In Intelligent Systems and Data Science, p. 19–32. External Links: ISBN 9789819796137, ISSN 1865-0937, Link, Document Cited by: §2.1. [35] L. Quach, K. N. Quoc, A. N. Quynh, and H. T. Ngoc (2022) Evaluation of the efficiency of the optimization algorithms for transfer learning on the rice leaf disease dataset. International Journal of Advanced Computer Science and Applications 13 (10). External Links: ISSN 2158-107X, Link, Document Cited by: §2.1. [36] K. N. Quoc, P. D. Dao, and L. Quach (2026) LeafNet: a large-scale dataset and comprehensive benchmark for foundational vision-language understanding of plant diseases. External Links: 2602.13662, Link Cited by: §2.2, §2.3. [37] K. N. Quoc, L. L. T. Thu, and L. Quach (2025) A vision-language foundation model for leaf disease identification. Expert Systems with Applications, p. 130084. Cited by: §2.2, §3.2, §4.2. [38] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. (2021) Learning transferable visual models from natural language supervision. In International conference on machine learning, p. 8748–8763. Cited by: §4.2. [39] S. N. Sakib, N. Haque, M. Z. Hossain, and S. E. Arman (2025) PlantVillageVQA: a visual question answering dataset for benchmarking vision-language models in plant science. arXiv. External Links: Document, Link Cited by: §2.3. [40] M. Shafay, D. Velayudhan, M. Owais, T. Hassan, L. Seneviratne, I. Hussain, and N. Werghi (2026) An end-to-end vision–language framework with llm-based reasoning for decision-oriented tomato leaf disease management in greenhouse environments. Computers and Electronics in Agriculture 247, p. 111692. External Links: ISSN 0168-1699, Link, Document Cited by: §2.3, §4.1, §4.1. [41] V. Solapure, S. DY, and A. JAWALE (2024) Tomato leaf disease dataset. Mendeley Data. External Links: Document, Link Cited by: §4.1, §4.1. [42] S. Stevens, J. Wu, M. J. Thompson, E. G. Campolongo, C. H. Song, D. E. Carlyn, L. Dong, W. M. Dahdul, C. Stewart, T. Berger-Wolf, et al. (2024) Bioclip: a vision foundation model for the tree of life. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 19412–19424. Cited by: §4.2. [43] C. Szegedy, S. Ioffe, V. Vanhoucke, and A. Alemi (2017) Inception-v4, inception-resnet and the impact of residual connections on learning. In Proceedings of the AAAI conference on artificial intelligence, Vol. 31. Cited by: §4.2. [44] M. Tan and Q. Le (2021) Efficientnetv2: smaller models and faster training. In International conference on machine learning, p. 10096–10106. Cited by: §4.2. [45] G. Truong, K. N. Quoc, and L. Quach (2026) TomaMMU: a comprehensive multimodal understanding benchmark for tomato leaf diseases. External Links: 2608.08727, Link Cited by: §3.4, §4.1, §4.1, §4.1. [46] M. Tschannen, A. Gritsenko, X. Wang, M. F. Naeem, I. Alabdulmohsin, N. Parthasarathy, T. Evans, L. Beyer, Y. Xia, B. Mustafa, et al. (2025) Siglip 2: multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786. Cited by: §4.2. [47] L. Wang, T. Jin, J. Yang, A. Leonardis, F. Wang, and F. Zheng (2024) Agri-llava: knowledge-infused large multimodal assistant on agricultural pests and diseases. arXiv preprint arXiv:2412.02158. Cited by: §6. [48] X. Wang and J. Liu (2024) An efficient deep learning model for tomato disease detection. Plant Methods 20 (1), p. 61. External Links: Document Cited by: §1, §2.1. [49] Y. Wen, Q. Li, Z. Ye, J. Zhang, X. Fan, Z. Mai, J. Wu, S. Lou, Y. Chen, H. Huang, Y. Zhang, D. Gu, L. Zhao, Y. Lu, H. Fu, J. Huang, and J. Zheng (2025) AgroCoT: a chain-of-thought benchmark for evaluating reasoning in vision-language models for agriculture. arXiv. External Links: Document, Link Cited by: §2.2. [50] H. Wu, B. Jia, and X. Yuan (2026) LLM-led vision-spectral fusion: a zero-shot approach to temporal fruit image classification. Neural Networks 194, p. 108155. External Links: ISSN 0893-6080, Link, Document Cited by: §2.1. [51] C. Xu, J. Ding, Y. Qiao, and L. Zhang (2022) Tomato disease and pest diagnosis method based on the stacking of prescription data. Computers and Electronics in Agriculture 197, p. 106997. External Links: Document Cited by: §1, §1. [52] Y. Zhao, S. Wang, Q. Zeng, W. Ni, H. Duan, N. Xie, and F. Xiao (2024) Informed-learning-guided visual question answering model of crop disease. Plant Phenomics 6, p. 0277. External Links: ISSN 2643-6515, Link, Document Cited by: §2.3. [53] H. Zhu, C. Lin, G. Liu, D. Wang, S. Qin, A. Li, J. Xu, and Y. He (2024) Intelligent agriculture: deep learning in uav-based remote sensing imagery for crop diseases and pests detection. Frontiers in Plant Science 15, p. 1435016. External Links: Document Cited by: §1. [54] H. Zhu, S. Qin, M. Su, C. Lin, A. Li, and J. Gao (2025) Harnessing large vision and language models in agriculture: a review. Frontiers in Plant Science 16. External Links: ISSN 1664-462X, Link, Document Cited by: §2.2. [55] J. Zhu, W. Wang, Z. Chen, Z. Liu, S. Ye, L. Gu, H. Tian, Y. Duan, W. Su, J. Shao, et al. (2025) Internvl3: exploring advanced training and test-time recipes for open-source multimodal models. arXiv preprint arXiv:2504.10479. Cited by: §4.2, Table 1, Table 1. [56] G. B. Zimmermann, M. E. Pellenz, Y. M. G. Costa, and A. d. S. Britto Jr. (2025) Enhancing disease and pest detection in greenhouse tomato cultivation using advanced machine learning on new dataset of images. Journal of the Brazilian Computer Society 31 (1), p. 187–202. External Links: ISSN 1678-4804, Link, Document Cited by: §4.1, §4.1. Appendix A Appendix Section Figure 1: Class distribution of the dataset across the six VQA tasks. (a) For the SI task, we show the top 15 classes with the highest image counts, where abbreviations are: Yell (yellowing), Curl (curling), DBMS (dark brown margin spot), TGCS (tan to gray center spot), WSLe (water-soaked lesion), CRin (concentric ring), SSpo (small spot), ISha (irregularly shaped), RBSp (reddish-brown spot), BPYe (blotched with pale yellow), NLes (necrotic lesion), Mott (mottling), YMos (yellow mosaic), DYLe (distortion of younger leaves), and SDLe (small dark lesion). (b) For the SNC task, abbreviations are: TYLC (Tomato yellow leaf curl virus), SLyc (Septoria lycopersici), TMVi (Tomato mosaic virus), ALin (Alternaria linariae), XVes (Xanthomonas vesicatoria), TUrt (Tetranychus urticae), PInf (Phytophthora infestans), Ccas (Corynespora cassiicola), PFul (Passalora fulva), Valb (Verticillium albo-atrum), Rsol (Ralstonia solanacearum), and Foxy (Fusarium oxysporum). (c) For the HDC task, classes are Disease and Healthy. (d) For the PC task, classes are Fungal, Virus, Bacterial, and Spider. (e) For the LC task, classes are One and Many. (f) For the DC task, abbreviations are: YLCV (Yellow Leaf Curl Virus), SLSp (Septoria Leaf Spot), EBli (Early Blight), BSpo (Bacterial Spot), LBli (Late Blight), SMTS (Spider mites Two-spotted Spider Mite), Viro (Virosis), TSpo (Target Spot), LMol (Leaf Mold), MVir (Mosaic Virus), VWil (Verticillium Wilt), Mite (Mite), BFlo (Bacterial Floundering), and FWil (Fusarium Wilt). SYSTEM_PROMPT You are an expert multimodal AI evaluator. Your task is to evaluate the accuracy of a candidate answer to a visual question, based on the provided ground truth answer set. PROMPT_TEMPLATE Ground Truth Reference Answers: ground_truth_answers Prediction Answer: prediction_answer Evaluate the candidate answer based on the following scale (1 to 5): 1. The answer is completely incorrect, hallucinatory, or irrelevant to the question. 2. The answer is partially relevant but contains major inaccuracies or misses the core question. 3. The answer is generally correct but lacks specific details, or contains minor factual errors. 4. The answer is correct, directly addresses the question, and is semantically equivalent to the ground truth. 5. The answer is exceptionally accurate, clear, and demonstrates a strong understanding of the visual and textual context. Respond with a score (1, 2, 3, 4, or 5). Figure 2: Prompt template used for LLM-as-a-Judge evaluation of visual question answering (VQA) responses on a five-point scale. The system prompt (top) specifies the evaluator role and behavioral constraints; the user prompt template (bottom) supplies input placeholders (ground_truth_answers and prediction_answer), the 5-point evaluation criteria, and restricts the output strictly to a single integer for automated parsing. Table 1: Performance comparison of SOLAR against foundation models, vision models, and fine-tuned vision-language models across four tomato disease VQA datasets. Model TomaMMU TLID TLD-3 TomaAD Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Foundation Models InternVL3 1B 55.21 0.552 27.04 0.270 53.80 0.538 50.20 0.441 46.56 0.450 InternVL3 2B 56.20 0.562 38.49 0.385 55.01 0.550 59.99 0.541 52.42 0.510 Qwen2.5 3B 58.00 0.580 38.33 0.383 62.46 0.582 55.62 0.503 53.60 0.512 Qwen3 2B 58.87 0.589 33.02 0.330 58.01 0.579 43.87 0.389 48.44 0.472 LLaVA-OV 52.42 0.524 33.04 0.330 53.80 0.538 42.63 0.314 45.47 0.427 LFM-VL 2.5 58.72 0.587 32.03 0.320 56.89 0.569 50.33 0.445 49.49 0.480 SCOLD 37.19 0.372 37.43 0.374 38.81 0.388 35.11 0.335 37.14 0.367 BioCLIP 48.95 0.567 22.05 0.221 50.63 0.506 28.42 0.270 37.51 0.391 Vision Models ResNet50 86.40 0.855 32.27 0.327 83.79 0.812 31.45 0.314 58.48 0.577 EfficientNetV2S 85.91 0.849 37.78 0.428 82.72 0.833 36.54 0.426 60.74 0.634 DenseNet121 86.15 0.853 31.99 0.323 80.31 0.783 32.57 0.360 57.75 0.580 InceptionResNetV2 84.74 0.836 25.44 0.290 82.81 0.807 29.26 0.313 55.56 0.561 Fine-tuned VLMs CLIP 93.83 0.946 21.91 0.218 85.07 0.799 26.42 0.261 56.81 0.556 BioCLIP 89.61 0.911 29.23 0.292 82.61 0.778 30.08 0.299 57.88 0.570 SigLIP2 91.49 0.921 17.59 0.173 87.62 0.862 37.20 0.361 58.47 0.579 AgriCLIP 93.73 0.948 25.04 0.249 35.68 0.350 23.63 0.233 44.52 0.445 SCOLD 83.70 0.733 30.86 0.306 48.86 0.440 28.92 0.255 48.09 0.433 LLaVA 96.28 0.944 30.02 0.322 91.29 0.873 50.11 0.403 66.92 0.636 SOLAR 97.68 0.974 33.70 0.335 91.67 0.855 53.20 0.501 69.06 0.666 Table 2: Performance comparison of SOLAR against foundation models, vision models, and fine-tuned vision-language models. The values represent the average results across four tomato disease VQA datasets categorized by question-based tasks. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Foundation Models InternVL3 1B 63.76 0.550 24.71 0.247 34.85 0.349 37.38 0.373 28.25 0.280 90.45 0.904 46.56 0.450 InternVL3 2B 88.25 0.798 39.22 0.391 32.16 0.321 39.36 0.393 25.34 0.253 90.22 0.901 52.42 0.510 Qwen2.5 3B 86.41 0.817 36.80 0.358 37.52 0.373 42.71 0.361 29.86 0.298 88.32 0.867 53.60 0.512 Qwen3 2B 86.23 0.789 27.57 0.275 17.53 0.175 38.42 0.384 31.66 0.318 89.23 0.890 48.44 0.472 LLaVA-OV 75.54 0.668 26.10 0.243 24.49 0.210 32.12 0.310 27.44 0.262 87.17 0.868 45.47 0.427 LFM-VL 2.5 82.74 0.808 22.00 0.220 26.27 0.258 41.26 0.406 35.15 0.351 89.54 0.840 49.49 0.480 SCOLD 44.80 0.448 44.80 0.447 42.51 0.425 35.25 0.352 34.97 0.350 20.48 0.182 37.14 0.367 BioCLIP 62.90 0.688 22.60 0.185 23.74 0.282 29.39 0.324 22.33 0.207 64.12 0.661 37.51 0.391 Vision Models ResNet50 93.25 0.950 44.42 0.450 46.77 0.486 53.01 0.554 23.63 0.222 89.79 0.800 58.48 0.577 EfficientNetV2S 90.15 0.929 48.85 0.529 50.76 0.541 62.05 0.660 26.24 0.291 86.37 0.855 60.74 0.634 DenseNet121 93.06 0.950 45.01 0.460 46.11 0.477 60.08 0.614 16.03 0.154 86.26 0.826 57.75 0.580 InceptionResNetV2 80.97 0.877 43.06 0.435 43.90 0.449 54.22 0.564 22.96 0.225 88.28 0.821 55.56 0.561 Fine-tuned VLMs CLIP 79.21 0.792 49.17 0.470 51.53 0.515 56.09 0.561 39.88 0.348 64.97 0.650 56.81 0.556 BioCLIP 82.82 0.828 44.37 0.425 46.44 0.464 61.31 0.613 34.78 0.315 77.57 0.776 57.88 0.570 SigLIP2 74.37 0.744 49.57 0.475 59.01 0.590 62.97 0.630 46.82 0.456 58.13 0.581 58.47 0.579 AgriCLIP 80.29 0.803 46.08 0.439 50.24 0.502 53.90 0.539 41.20 0.364 70.40 0.704 44.52 0.445 SCOLD 82.42 0.761 43.37 0.420 54.66 0.494 48.72 0.488 15.44 0.089 82.30 0.782 48.09 0.433 LLaVA 89.14 0.881 52.84 0.529 59.49 0.553 63.24 0.575 48.92 0.424 87.93 0.851 66.92 0.636 SOLAR 93.61 0.936 57.44 0.552 61.78 0.618 64.04 0.641 53.55 0.411 83.98 0.840 69.06 0.666 Table 3: Performance comparison of different components in the SOLAR architecture. The results show that SOLAR outperforms other configurations, showing that all components are important. TomaMMU TLID TLD-3 TomaAD Average Model Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 W/oW/o Fusion Adapter 97.20 0.969 34.05 0.337 92.03 0.859 51.65 0.490 68.73 0.664 W/oW/o MoE 98.04 0.977 33.44 0.330 90.56 0.846 42.75 0.416 66.20 0.642 SOLAR 97.68 0.974 33.70 0.335 91.67 0.855 53.20 0.501 69.06 0.666 Table 4: Performance comparison of different vision encoders across four datasets in the SOLAR architecture TomaMMU TLID TLD-3 TomaAD Average Model Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 CLIP 96.53 0.958 38.17 0.373 86.80 0.822 40.95 0.403 65.61 0.639 BioCLIP 97.21 0.976 32.96 0.329 85.21 0.852 35.65 0.356 62.76 0.628 AgriCLIP 96.94 0.967 46.13 0.451 85.09 0.798 31.99 0.318 65.04 0.634 SCOLD 97.68 0.974 33.70 0.335 91.67 0.855 53.20 0.501 69.06 0.666 Table 5: Performance comparison of different LLM backbones across four datasets in the SOLAR architecture TomaMMU TLID TLD-3 TomaAD Average Model Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Qwen 3 97.24 0.978 34.25 0.343 89.42 0.855 47.38 0.454 67.07 0.658 BioGPT 97.76 0.974 32.00 0.318 92.09 0.861 48.20 0.452 67.51 0.651 LFM-VL 2.5 97.68 0.974 33.70 0.335 91.67 0.855 53.20 0.501 69.06 0.666 Table 6: Performance comparison across six VQA tasks on the TomaMMU dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Foundation Models InternVL3 1B 89.88 0.899 28.61 0.286 30.69 0.307 50.82 0.508 37.54 0.375 93.73 0.937 55.21 0.552 InternVL3 2B 82.08 0.821 34.23 0.342 42.65 0.426 51.44 0.514 32.42 0.324 94.38 0.944 56.20 0.562 Qwen2.5 3B 83.84 0.838 40.60 0.406 47.46 0.475 51.01 0.510 32.46 0.325 92.64 0.926 58.00 0.580 Qwen3 2B 91.55 0.915 45.55 0.456 32.63 0.326 55.88 0.559 33.94 0.339 93.68 0.937 58.87 0.589 LLaVA-OV 89.88 0.899 34.75 0.347 24.82 0.248 45.16 0.452 26.69 0.267 93.23 0.932 52.42 0.524 LFM-VL 2.5 90.92 0.909 41.94 0.419 38.07 0.381 54.05 0.540 33.02 0.330 94.29 0.943 58.72 0.587 SCOLD 25.02 0.250 63.43 0.634 43.40 0.434 40.60 0.406 39.46 0.395 11.25 0.113 37.19 0.372 BioCLIP 63.12 0.867 46.55 0.300 35.50 0.533 47.73 0.600 20.02 0.133 80.78 0.967 48.95 0.567 Vision Models ResNet50 98.68 0.993 89.98 0.900 90.42 0.903 96.24 0.962 43.90 0.406 99.17 0.962 86.40 0.855 EfficientNetV2S 98.20 0.990 88.87 0.887 89.86 0.898 95.98 0.960 43.76 0.416 98.76 0.945 85.91 0.849 DenseNet121 98.70 0.993 89.79 0.896 90.04 0.899 96.02 0.960 43.17 0.406 99.17 0.963 86.15 0.853 InceptionResNetV2 98.02 0.989 85.30 0.845 86.61 0.862 95.98 0.960 43.76 0.416 98.76 0.945 84.74 0.836 Fine-tuned VLMs CLIP 96.23 0.962 94.92 0.949 94.48 0.945 98.38 0.984 81.27 0.858 97.72 0.977 93.83 0.946 BioCLIP 97.69 0.977 89.07 0.890 88.89 0.889 95.94 0.959 68.12 0.771 97.93 0.979 89.61 0.911 SigLIP2 96.22 0.962 91.67 0.917 90.21 0.902 96.45 0.965 76.78 0.802 97.62 0.976 91.49 0.921 AgriCLIP 98.23 0.982 94.90 0.949 95.13 0.951 98.26 0.983 77.42 0.840 98.44 0.984 93.73 0.948 SCOLD 88.33 0.782 85.71 0.779 85.42 0.767 92.08 0.889 54.40 0.266 96.27 0.916 83.70 0.733 LLaVA 98.89 0.982 94.26 0.931 94.65 0.930 99.11 0.990 91.53 0.851 99.25 0.981 96.28 0.944 SOLAR 98.75 0.988 98.24 0.982 98.40 0.984 99.50 0.995 92.15 0.903 99.07 0.991 97.68 0.974 Table 7: Performance comparison across six VQA tasks on the TLID dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Foundation Models InternVL3 1B 0.00 0.000 0.00 0.000 42.67 0.427 1.84 0.018 17.97 0.180 99.78 0.998 27.04 0.270 InternVL3 2B 95.44 0.954 25.85 0.259 4.30 0.043 0.04 0.000 5.34 0.053 99.98 1.000 38.49 0.385 Qwen2.5 3B 93.04 0.930 27.35 0.274 2.63 0.026 0.21 0.002 6.98 0.070 99.78 0.998 38.33 0.383 Qwen3 2B 81.41 0.814 10.44 0.104 0.71 0.007 0.08 0.001 5.57 0.056 99.94 0.999 33.02 0.330 LLaVA-OV 47.12 0.471 14.95 0.149 24.30 0.243 0.00 0.000 12.01 0.120 99.88 0.999 33.04 0.330 LFM-VL 2.5 74.58 0.746 5.05 0.051 2.38 0.024 0.00 0.000 10.26 0.103 99.94 0.999 32.03 0.320 SigLIP2 25.65 0.257 29.19 0.292 34.41 0.344 14.95 0.149 3.96 0.040 10.47 0.105 19.77 0.198 SCOLD 76.20 0.762 27.27 0.273 22.76 0.228 51.94 0.519 46.37 0.464 0.02 0.000 37.43 0.374 BioCLIP 83.08 0.831 2.67 0.027 1.25 0.013 2.21 0.022 4.69 0.047 38.38 0.384 22.05 0.221 Vision Models ResNet50 90.99 0.911 0.29 0.006 0.58 0.012 0.04 0.001 1.72 0.033 100.00 1.000 32.27 0.327 EfficientNetV2S 91.28 0.912 9.85 0.179 2.34 0.046 12.90 0.229 13.03 0.214 97.29 0.986 37.78 0.428 DenseNet121 90.26 0.904 0.29 0.006 0.88 0.017 0.13 0.003 0.40 0.008 100.00 1.000 31.99 0.323 InceptionResNetV2 47.79 0.643 1.34 0.026 1.13 0.022 0.00 0.000 2.40 0.046 100.00 1.000 25.44 0.290 Fine-tuned VLMs CLIP 91.15 0.912 0.33 0.003 0.50 0.005 0.50 0.005 2.37 0.020 36.61 0.366 21.91 0.218 BioCLIP 92.54 0.925 0.00 0.000 0.04 0.000 0.08 0.001 0.73 0.005 82.00 0.820 29.23 0.292 SigLIP2 81.84 0.818 1.21 0.012 1.50 0.015 4.68 0.047 7.60 0.056 8.72 0.087 17.59 0.173 AgriCLIP 86.07 0.861 0.04 0.000 0.04 0.000 0.50 0.005 8.08 0.072 55.50 0.555 25.04 0.249 SCOLD 78.06 0.763 1.17 0.012 1.29 0.013 5.39 0.054 0.08 0.002 99.17 0.992 30.86 0.306 LLaVA 76.23 0.856 0.17 0.003 0.58 0.012 0.17 0.003 3.33 0.058 99.63 0.998 30.02 0.322 SOLAR 95.18 0.952 0.79 0.008 0.67 0.007 2.17 0.022 4.49 0.032 98.92 0.989 33.70 0.335 Table 8: Performance comparison across six VQA tasks on the TLD-3 dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Foundation Models InternVL3 1B 85.89 0.859 27.25 0.273 20.07 0.201 67.46 0.675 25.74 0.257 96.41 0.964 53.80 0.538 InternVL3 2B 84.83 0.848 23.16 0.232 31.70 0.317 68.01 0.680 26.47 0.265 95.91 0.959 55.01 0.550 Qwen2.5 3B 85.73 0.857 36.16 0.362 38.53 0.385 95.18 0.697 23.96 0.240 95.18 0.952 62.46 0.582 Qwen3 2B 91.29 0.912 34.20 0.336 32.52 0.325 69.58 0.696 24.22 0.242 96.22 0.962 58.01 0.579 LLaVA-OV 85.87 0.858 35.59 0.356 25.22 0.252 56.93 0.569 25.19 0.252 94.00 0.940 53.80 0.538 LFM-VL 2.5 87.57 0.876 31.96 0.320 32.45 0.325 70.79 0.708 22.21 0.222 96.33 0.963 56.89 0.569 SigLIP2 32.63 0.326 15.76 0.158 28.13 0.281 69.99 0.700 17.90 0.179 10.56 0.106 29.16 0.292 SCOLD 42.30 0.423 49.55 0.495 36.28 0.363 38.39 0.384 48.67 0.487 17.64 0.176 38.81 0.388 BioCLIP 66.74 0.667 39.92 0.399 30.56 0.306 62.84 0.628 17.91 0.179 85.80 0.858 50.63 0.506 Vision Models ResNet50 96.88 0.982 87.16 0.889 86.75 0.871 93.17 0.936 43.26 0.400 95.52 0.796 83.79 0.812 EfficientNetV2S 96.45 0.979 87.84 0.893 88.18 0.885 91.97 0.925 36.92 0.361 94.94 0.953 82.72 0.833 DenseNet121 96.24 0.979 87.92 0.897 86.71 0.870 95.30 0.957 20.53 0.202 95.15 0.795 80.31 0.783 InceptionResNetV2 96.02 0.977 83.57 0.849 84.06 0.843 93.93 0.943 43.19 0.417 96.08 0.815 82.81 0.807 Fine-tuned VLMs CLIP 91.93 0.919 85.57 0.770 85.43 0.854 91.97 0.920 70.06 0.474 85.44 0.854 85.07 0.799 BioCLIP 93.94 0.939 81.36 0.738 81.19 0.812 87.76 0.878 67.57 0.465 83.84 0.838 82.61 0.778 SigLIP2 91.66 0.917 86.00 0.775 86.14 0.861 89.02 0.890 86.45 0.865 86.45 0.865 87.62 0.862 AgriCLIP 93.94 0.939 84.85 0.761 85.38 0.854 91.66 0.917 73.03 0.502 85.15 0.852 85.67 0.804 SCOLD 92.82 0.877 78.80 0.793 88.86 0.912 87.59 0.890 6.43 0.062 92.24 0.834 74.46 0.728 LLaVA 97.00 0.941 89.72 0.934 90.35 0.932 97.26 0.988 78.10 0.565 95.33 0.877 91.29 0.873 SOLAR 96.86 0.969 91.19 0.823 91.35 0.913 96.79 0.968 79.22 0.510 94.62 0.946 91.67 0.855 Table 9: Performance comparison across six VQA tasks on the TomaAD dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Foundation Models InternVL3 1B 79.28 0.442 42.96 0.429 45.98 0.459 29.40 0.290 31.74 0.306 71.86 0.718 50.20 0.441 InternVL3 2B 90.64 0.570 73.62 0.732 50.00 0.499 37.94 0.377 37.14 0.368 70.60 0.701 59.99 0.541 Qwen2.5 3B 83.03 0.641 43.07 0.390 61.46 0.605 24.43 0.235 56.04 0.557 65.67 0.592 55.62 0.503 Qwen3 2B 80.68 0.516 20.10 0.202 4.27 0.043 28.14 0.280 62.92 0.633 67.09 0.660 43.87 0.389 LLaVA-OV 79.28 0.442 19.10 0.120 23.62 0.096 26.38 0.218 45.85 0.410 61.56 0.600 42.63 0.314 LFM-VL 2.5 77.89 0.700 9.05 0.089 32.16 0.303 40.20 0.375 75.10 0.750 67.59 0.453 50.33 0.445 SigLIP2 22.91 0.204 8.79 0.088 16.58 0.166 26.38 0.255 0.21 0.190 50.50 0.349 20.90 0.177 SCOLD 35.66 0.356 38.94 0.384 67.59 0.675 10.05 0.100 5.39 0.054 53.02 0.438 35.11 0.335 BioCLIP 38.65 0.386 1.26 0.013 27.64 0.275 4.77 0.047 46.68 0.467 51.51 0.434 28.42 0.270 Vision Models ResNet50 86.43 0.914 0.25 0.005 9.32 0.158 22.60 0.316 5.62 0.050 64.47 0.440 31.45 0.314 EfficientNetV2S 74.65 0.833 8.82 0.156 22.67 0.336 47.36 0.524 11.25 0.172 54.49 0.537 36.54 0.426 DenseNet121 87.03 0.922 2.02 0.040 6.80 0.120 48.87 0.534 0.00 0.000 50.70 0.545 32.57 0.360 InceptionResNetV2 82.04 0.897 2.02 0.019 3.78 0.068 26.95 0.351 2.50 0.020 58.28 0.522 29.26 0.313 Fine-tuned VLMs CLIP 37.52 0.375 15.87 0.159 25.69 0.257 33.50 0.335 5.83 0.038 40.12 0.401 26.42 0.261 BioCLIP 47.11 0.471 7.05 0.070 15.62 0.156 61.46 0.615 2.71 0.020 46.51 0.465 30.08 0.299 SigLIP2 27.74 0.277 19.40 0.194 58.19 0.582 61.71 0.617 16.46 0.100 39.72 0.397 37.20 0.361 AgriCLIP 42.91 0.429 4.53 0.045 20.40 0.204 25.19 0.252 6.25 0.043 42.51 0.425 23.63 0.233 SCOLD 70.46 0.621 7.81 0.094 43.07 0.282 9.82 0.120 0.83 0.026 41.52 0.385 28.92 0.255 LLaVA 84.43 0.746 27.20 0.246 52.39 0.337 56.42 0.319 22.71 0.220 57.49 0.548 50.11 0.403 SOLAR 83.63 0.836 39.55 0.395 56.68 0.567 57.68 0.577 38.33 0.197 43.31 0.433 53.20 0.501 Table 10: Performance comparison (ROUGELROUGE_L and GPTGPT) across six VQA tasks on the TomaMMU dataset. Model HDC DC SNC PC SI LC Average ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT InternVL3 1B 0.987 4.595 0.871 2.474 0.840 2.230 0.965 3.033 0.897 2.782 0.993 4.749 0.926 3.311 InternVL3 2B 0.978 4.285 0.883 2.739 0.876 2.726 0.965 3.058 0.889 2.634 0.994 4.775 0.931 3.370 Qwen2.5 3B 0.980 4.354 0.875 2.630 0.885 2.937 0.965 3.041 0.888 2.620 0.992 4.709 0.931 3.382 Qwen3 2B 0.990 4.684 0.895 2.923 0.852 2.311 0.969 3.235 0.891 2.676 0.993 4.747 0.932 3.429 LLaVA-OV 0.987 4.595 0.866 2.413 0.837 2.014 0.961 2.807 0.883 2.358 0.993 4.729 0.921 3.153 LFM-VL 2.5 0.989 4.637 0.905 3.089 0.862 2.530 0.967 3.162 0.891 2.621 0.994 4.789 0.935 3.471 SOLAR 0.998 4.054 0.997 4.858 0.997 4.936 1.000 4.980 0.976 4.653 0.999 4.963 0.995 4.741 Table 11: Performance comparison (ROUGELROUGE_L and GPTGPT) across six VQA tasks on the TLID dataset. Model HDC DC SNC PC SI LC Average ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT InternVL3 1B 0.996 4.877 0.819 1.634 0.841 1.005 0.927 1.042 0.872 1.377 0.995 4.816 0.908 2.459 InternVL3 2B 0.995 4.844 0.890 2.155 0.821 1.712 0.962 3.235 0.847 3.008 0.999 4.980 0.919 3.319 Qwen2.5 3B 0.993 4.767 0.873 2.064 0.803 1.890 0.929 1.000 0.862 1.980 0.994 4.798 0.909 2.750 Qwen3 2B 0.995 4.829 0.917 2.393 0.832 1.096 0.976 3.777 0.865 1.742 0.995 4.828 0.930 3.111 LLaVA-OV 0.941 3.116 0.891 1.896 0.838 1.901 0.929 1.027 0.852 2.194 0.989 5.000 0.906 2.522 LFM-VL 2.5 0.941 3.116 0.822 2.006 0.828 1.060 0.926 1.095 0.839 1.973 0.949 4.381 0.884 2.272 SOLAR 0.994 4.807 0.842 1.581 0.825 1.027 0.930 1.087 0.772 2.433 0.999 4.957 0.894 2.649 Table 12: Performance comparison (ROUGELROUGE_L and GPTGPT) across six VQA tasks on the TLD-3 dataset. Model HDC DC SNC PC SI LC Average ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT InternVL3 1B 0.987 4.595 0.871 2.474 0.840 2.230 0.965 3.033 0.897 2.782 0.993 4.749 0.926 3.311 InternVL3 2B 0.978 4.285 0.883 2.739 0.876 2.726 0.965 3.058 0.889 2.634 0.994 4.775 0.931 3.370 Qwen2.5 3B 0.980 4.354 0.875 2.630 0.885 2.937 0.965 3.041 0.888 2.620 0.992 4.709 0.931 3.382 Qwen3 2B 0.990 4.684 0.895 2.923 0.852 2.311 0.969 3.235 0.891 2.676 0.993 4.747 0.932 3.429 LLaVA-OV 0.987 4.595 0.866 2.413 0.837 2.014 0.961 2.807 0.883 2.358 0.993 4.729 0.921 3.153 LFM-VL 2.5 0.989 4.637 0.905 3.089 0.862 2.530 0.967 3.162 0.891 2.621 0.994 4.789 0.935 3.471 SOLAR 0.996 4.874 0.979 4.596 0.985 4.658 0.998 4.872 0.886 3.511 0.994 4.785 0.973 4.549 Table 13: Performance comparison (ROUGELROUGE_L and GPTGPT) across six VQA tasks on the TomaAD dataset. Model HDC DC SNC PC SI LC Average ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT ROUGELROUGE_L GPTGPT InternVL3 1B 0.985 4.505 0.825 1.524 0.791 1.249 0.928 1.771 0.886 1.785 0.915 2.014 0.888 2.141 InternVL3 2B 0.988 4.609 0.823 1.947 0.783 1.191 0.940 1.741 0.868 2.710 0.939 2.900 0.890 2.516 Qwen2.5 3B 0.981 4.385 0.804 1.794 0.769 1.106 0.930 1.337 0.876 2.265 0.932 2.625 0.882 2.252 Qwen3 2B 0.982 4.425 0.859 1.723 0.798 1.287 0.927 2.096 0.875 2.013 0.915 2.062 0.893 2.268 LLaVA-OV 0.901 1.830 0.842 2.063 0.791 1.806 0.936 1.685 0.880 1.973 0.961 4.723 0.885 2.347 LFM-VL 2.5 0.901 1.827 0.823 1.559 0.806 1.642 0.909 1.356 0.816 1.994 0.915 2.634 0.862 1.835 SOLAR 0.980 4.345 0.903 3.098 0.911 3.282 0.970 3.390 0.817 3.156 0.937 2.732 0.920 3.334 Table 14: Performance comparison of different vision encoders across six VQA tasks on the TomaMMU dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 CLIP 98.59 0.986 96.97 0.970 96.45 0.965 98.73 0.987 89.35 0.849 99.11 0.991 96.53 0.958 BioCLIP 98.59 0.986 98.01 0.980 97.08 0.971 98.76 0.988 91.61 0.940 99.19 0.992 97.21 0.976 AgriCLIP 98.68 0.987 98.01 0.979 96.16 0.962 98.67 0.987 90.86 0.897 99.23 0.992 96.94 0.967 SCOLD 98.75 0.988 98.24 0.982 98.40 0.984 99.50 0.995 92.15 0.903 99.07 0.991 97.68 0.974 Table 15: Performance comparison of different vision encoders across six VQA tasks on the TLID dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 CLIP 93.74 0.937 8.73 0.087 7.85 0.079 4.05 0.041 14.69 0.094 99.94 0.999 38.17 0.373 BioCLIP 96.10 0.961 0.08 0.001 0.04 0.000 0.92 0.009 3.53 0.035 97.07 0.971 32.96 0.329 AgriCLIP 81.68 0.817 16.41 0.164 29.81 0.298 30.44 0.304 19.33 0.131 99.13 0.991 46.13 0.451 SCOLD 95.18 0.952 0.79 0.008 0.67 0.007 2.17 0.022 4.49 0.032 98.92 0.989 33.70 0.335 Table 16: Performance comparison of different vision encoders across six VQA tasks on the TLD-3 dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 CLIP 96.10 0.961 84.01 0.840 81.97 0.820 91.68 0.917 71.01 0.435 96.00 0.960 86.80 0.822 BioCLIP 97.79 0.978 80.13 0.801 80.83 0.808 91.74 0.917 68.51 0.685 92.24 0.922 85.21 0.852 AgriCLIP 95.49 0.955 80.09 0.731 76.76 0.768 92.89 0.929 69.52 0.447 95.78 0.958 85.09 0.798 SCOLD 96.86 0.969 91.19 0.823 91.35 0.913 96.79 0.968 79.22 0.510 94.62 0.946 91.67 0.855 Table 17: Performance comparison of different vision encoders across six VQA tasks on the TomaAD dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 CLIP 84.43 0.844 20.15 0.201 17.38 0.174 69.52 0.695 7.92 0.043 46.31 0.463 40.95 0.403 BioCLIP 78.64 0.786 12.34 0.123 26.20 0.262 38.79 0.388 10.42 0.104 47.50 0.475 35.65 0.356 AgriCLIP 83.83 0.838 1.26 0.013 10.33 0.103 47.10 0.471 2.71 0.016 46.71 0.467 31.99 0.318 SCOLD 83.63 0.836 39.55 0.395 56.68 0.567 57.68 0.577 38.33 0.197 43.31 0.433 53.20 0.501 Table 18: Performance comparison of different LLM backbones across six VQA tasks on the TomaMMU dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Qwen 3 98.87 0.989 98.16 0.982 98.44 0.984 99.50 0.995 89.58 0.926 98.90 0.989 97.24 0.978 BioGPT 98.96 0.990 97.53 0.975 98.16 0.982 99.54 0.995 93.52 0.913 98.86 0.989 97.76 0.974 LFM-VL 2.5 98.75 0.988 98.24 0.982 98.40 0.984 99.50 0.995 92.15 0.903 99.07 0.991 97.68 0.974 Table 19: Performance comparison of different LLM backbones across six VQA tasks on the TLID dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Qwen 3 95.38 0.954 0.88 0.009 1.67 0.017 2.25 0.022 6.16 0.062 99.17 0.992 34.25 0.343 BioGPT 95.59 0.956 0.58 0.006 0.63 0.006 0.42 0.004 4.78 0.037 90.01 0.900 32.00 0.318 LFM-VL 2.5 95.18 0.952 0.79 0.008 0.67 0.007 2.17 0.022 4.49 0.032 98.92 0.989 33.70 0.335 Table 20: Performance comparison of different LLM backbones across six VQA tasks on the TLD-3 dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Qwen 3 97.72 0.977 81.97 0.820 91.64 0.916 97.38 0.974 72.80 0.495 95.04 0.95 89.42 0.855 BioGPT 98.28 0.983 91.97 0.832 91.52 0.915 96.97 0.970 78.50 0.511 95.32 0.953 92.09 0.861 LFM-VL 2.5 96.86 0.969 91.19 0.823 91.35 0.913 96.79 0.968 79.22 0.510 94.62 0.946 91.67 0.855 Table 21: Performance comparison of different LLM backbones across six VQA tasks on the TomaAD dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Qwen 3 84.23 0.842 28.72 0.287 37.53 0.375 59.45 0.595 26.25 0.143 48.10 0.481 47.38 0.454 BioGPT 82.83 0.828 29.72 0.297 36.02 0.360 55.92 0.559 35.62 0.179 49.10 0.491 48.20 0.452 LFM-VL 2.5 83.63 0.836 39.55 0.395 56.68 0.567 57.68 0.577 38.33 0.197 43.31 0.433 53.20 0.501 Table 22: Performance comparison of different MoE placement strategies across six VQA tasks on the TomaMMU dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Dual MoE 98.96 0.990 98.61 0.986 98.78 0.988 99.54 0.995 94.37 0.926 99.21 0.992 98.25 0.979 Fusion MoE 98.91 0.989 96.70 0.957 97.26 0.973 99.30 0.993 91.51 0.902 98.44 0.984 97.02 0.966 Image MoE 98.75 0.988 98.24 0.982 98.40 0.984 99.50 0.995 92.15 0.903 99.07 0.991 97.68 0.974 Table 23: Performance comparison of different MoE placement strategies across six VQA tasks on the TLID dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Dual MoE 93.13 0.931 0.42 0.004 0.25 0.003 0.75 0.007 3.65 0.029 98.76 0.988 32.83 0.327 Fusion MoE 94.53 0.942 0.50 0.005 0.21 0.002 0.79 0.008 0.88 0.007 99.88 0.993 32.80 0.326 Image MoE 95.18 0.952 0.79 0.008 0.67 0.007 2.17 0.022 4.49 0.032 98.92 0.989 33.70 0.335 Table 24: Performance comparison of different MoE placement strategies across six VQA tasks on the TLD-3 dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Dual MoE 97.37 0.974 90.80 0.820 91.33 0.913 98.14 0.981 78.66 0.510 95.75 0.958 92.01 0.859 Fusion MoE 97.84 0.977 90.82 0.820 91.33 0.913 97.47 0.975 77.18 0.496 94.96 0.949 91.60 0.855 Image MoE 96.86 0.969 91.19 0.823 91.35 0.913 96.79 0.968 79.22 0.510 94.62 0.946 91.67 0.855 Table 25: Performance comparison of different MoE placement strategies across six VQA tasks on the TomaAD dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Dual MoE 83.63 0.836 17.13 0.171 53.15 0.531 47.86 0.479 16.04 0.082 54.29 0.543 45.35 0.44 Fusion MoE 85.23 0.850 22.17 0.222 51.13 0.511 52.64 0.526 26.88 0.137 58.28 0.581 49.39 0.471 Image MoE 83.63 0.836 39.55 0.395 56.68 0.567 57.68 0.577 38.33 0.197 43.31 0.433 53.20 0.501 Table 26: Performance comparison of different MoE placement strategies across six VQA tasks on the TomaMMU dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 W/oW/o Fusion Adapter 98.87 0.989 96.64 0.966 96.72 0.967 99.32 0.993 92.62 0.909 99.03 0.990 97.20 0.969 W/oW/o MoE 99.01 0.990 98.42 0.984 98.47 0.985 99.54 0.995 93.78 0.919 99.01 0.990 98.04 0.977 SOLAR 98.75 0.988 98.24 0.982 98.40 0.984 99.50 0.995 92.15 0.903 99.07 0.991 97.68 0.974 Table 27: Performance comparison of different MoE placement strategies across six VQA tasks on the TLID dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 W/oW/o Fusion Adapter 94.73 0.947 0.25 0.003 0.17 0.002 0.25 0.003 8.96 0.067 99.92 0.998 34.05 0.337 W/oW/o MoE 91.72 0.917 0.42 0.004 0.71 0.007 0.50 0.005 7.35 0.048 99.94 0.999 33.44 0.330 SOLAR 95.18 0.952 0.79 0.008 0.67 0.007 2.17 0.022 4.49 0.032 98.92 0.989 33.70 0.335 Table 28: Performance comparison of different MoE placement strategies across six VQA tasks on the TLD-3 dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 W/oW/o Fusion Adapter 98.17 0.982 91.27 0.824 91.80 0.918 97.22 0.972 78.53 0.507 95.21 0.952 92.03 0.859 W/oW/o MoE 97.47 0.975 89.57 0.808 88.82 0.888 95.01 0.950 77.45 0.507 95.05 0.951 90.56 0.846 SOLAR 96.86 0.969 91.19 0.823 91.35 0.913 96.79 0.968 79.22 0.510 94.62 0.946 91.67 0.855 Table 29: Performance comparison of different MoE placement strategies across six VQA tasks on the TomaAD dataset. Model HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 W/oW/o Fusion Adapter 86.63 0.864 22.67 0.227 45.09 0.451 65.49 0.652 32.92 0.177 57.09 0.571 51.65 0.490 W/oW/o MoE 85.03 0.850 23.43 0.234 34.01 0.340 46.35 0.464 14.37 0.074 53.29 0.533 42.75 0.416 SOLAR 83.63 0.836 39.55 0.395 56.68 0.567 57.68 0.577 38.33 0.197 43.31 0.433 53.20 0.501 Table 30: Performance comparison of different MoE configurations (n and k) on the TomaMMU dataset. n k HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 4 2 98.75 98.75 98.24 98.24 98.40 98.40 99.50 99.50 92.15 90.27 99.07 99.07 97.68 97.37 6 1 98.78 98.78 96.39 96.39 96.66 96.66 98.69 98.69 91.07 89.16 99.25 99.25 96.81 96.49 6 2 98.92 98.60 95.77 97.30 97.39 98.20 99.36 99.60 92.79 88.10 99.17 98.80 97.23 96.77 8 2 98.91 98.91 97.55 97.55 97.10 97.10 99.03 99.03 92.58 91.45 99.23 99.23 97.40 97.21 Table 31: Performance comparison of different MoE configurations (n and k) on the TLID dataset. n k HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 4 2 95.18 95.18 0.79 0.79 0.67 0.67 2.17 2.17 4.49 3.17 98.92 98.92 33.70 33.48 6 1 47.12 47.12 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 7.85 7.85 6 2 96.48 98.20 4.80 9.20 2.71 5.30 0.58 1.20 4.72 9.40 96.10 98.00 34.23 36.88 8 2 96.54 96.54 0.25 0.25 0.33 0.33 0.33 0.33 1.19 0.89 99.53 99.53 33.03 32.98 Table 32: Performance comparison of different MoE configurations (n and k) on the TLD-3 dataset. n k HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 4 2 96.86 96.86 91.19 82.30 91.35 91.35 96.79 96.79 79.22 51.05 94.62 94.62 91.67 85.50 6 1 91.29 91.29 38.41 31.52 47.65 47.65 49.00 49.00 39.27 26.06 50.18 50.18 52.63 49.28 6 2 98.53 97.70 90.49 95.20 90.99 95.70 97.32 99.10 79.25 58.40 95.68 95.30 92.04 90.23 8 2 98.33 98.33 91.07 82.17 91.95 91.95 97.36 97.36 77.49 50.61 95.36 95.36 91.93 85.96 Table 33: Performance comparison of different MoE configurations (n and k) on the TomaAD dataset. n k HDC DC SNC PC SI LC Average Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 Acc F1 4 2 83.63 83.63 39.55 39.55 56.68 56.68 57.68 57.68 38.33 19.69 43.31 43.31 53.20 50.09 6 1 79.24 79.24 0.00 0.00 0.00 0.00 73.05 73.05 0.00 0.00 49.62 49.62 33.65 33.65 6 2 87.43 85.40 33.75 55.70 40.30 41.30 45.34 60.50 18.75 14.80 53.49 71.00 46.51 54.78 8 2 86.03 86.03 21.41 21.41 52.39 52.39 43.32 43.32 15.42 8.02 52.89 52.89 45.24 44.01