Social bots are increasingly polluting online platforms by spreading misinformation and engaging in coordinated manipulation, posing severe threats to cybersecurity. Graph Neural Networks (GNNs) have become mainstream for social bot detection due to their ability to integrate structural and attribute features, with spectral-based approaches demonstrating particular efficacy due to discriminative patterns in the spectral domain. However, current spectral GNN methods face two limitations: (1) their broad-spectrum fitting mechanisms degrade the focus on bot-specific spectral features, and (2) certain domain knowledge valuable for bot detection, e.g., low homophily correlates with high-frequency features, has not been fully incorporated into existing methods.

To address these challenges, we propose HWGNN, a novel homophily-aware graph spectral network with Gaussian window constraints. Our framework introduces two key innovations: (i) a Gaussian-window constrained spectral network that employs learnable Gaussian windows to highlight bot-related spectral features, and (ii) a homophily-aware adaptation mechanism that injects domain knowledge between homophily ratios and frequency features into the Gaussian window optimization process. Through extensive experimentation on multiple benchmark datasets, we demonstrate that HWGNN achieves state-of-the-art bot detection performance, outperforming existing methods with an average improvement of 4.6% in F1-score, while exhibiting strong plug-in compatibility with existing spectral GNNs.

Multimodal Table Question Answering requires both accurate perception of cell contents in a table image and precise symbolic reasoning over the extracted values; multimodal large language models (MLLMs) handle the former well, but the latter remains more challenging, motivating the use of an explicit code interpreter for computation. Through a fine-grained partition analysis, we find instance-level complementarity between the two paths: even when the visual path is stronger overall, the structured path correctly answers a non-trivial subset of questions missed by the visual path. To exploit this complementarity, we propose TriVTab, a dual-path framework that combines direct visual reasoning with triple-based structured reasoning through agreement-gated multimodal fusion. Rather than seeking a uniformly stronger standalone path, TriVTab reconciles complementary paths to improve a strong visual baseline. TriVTab (i) uses an execution-oriented cell-triple representation to bind each value to an explicit row and column context, making complex table structures directly addressable for code generation; and (ii) employs agreement-gated multimodal fusion to treat cross-path agreement as corroboration and concentrate multimodal fusion on divergent predictions. On five multimodal Table QA benchmarks, TriVTab improves over a strong Qwen3-VL-8B chain-of-thought baseline by +2.53 points on average and over a Self-Reflection baseline by +3.46 points.