Abstract: This report synthesises findings from 8 peer-reviewed papers addressing the following research question: What is the performance trade-off between single-view and multi-view aggregation methods in MECCH when evaluated on large-scale heterogeneous academic graphs (e.g., MAG240M) for node classification. Heterogeneous…

Abstract: This report synthesises findings from 8 peer-reviewed papers addressing the following research question: How do different metapath sampling strategies in MECCH compare to traditional random walk sampling in heterogeneous graph node classification benchmarks like ACM or DBLP in terms of classification. Heterogeneous…

Abstract: This report synthesises findings from 12 peer-reviewed papers addressing the following research question: What is the impact of ARS on the reasoning accuracy of multimodal LLMs (e.g., Flamingo or GIT) when evaluated on dynamic visual reasoning benchmarks such as VQA-v2 or TextVQA, compared to static. Recent advances…

Abstract: This report synthesises findings from 4 peer-reviewed papers addressing the following research question: How does the Rank VQA model's ranking-inspired hybrid training strategy compare to contrastive learning methods in terms of VQA accuracy on the VQA v2 and GQA benchmarks. Visual Question Answering (VQA) is a…

Abstract: This report synthesises findings from 9 peer-reviewed papers addressing the following research question: How do self-supervised contrastive learning approaches scale in terms of AUC-ROC performance when applied to cross-domain graph anomaly detection tasks with varying levels of heterophily. Graph Anomaly Detection…

Abstract: This report synthesises findings from 11 peer-reviewed papers addressing the following research question: How does the scalability of ARS compare to reinforcement learning-based reasoning suppression methods when applied to larger models (e.g., >100B parameters) in terms of throughput and accuracy on the. Large…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: What is the impact of incorporating multimodal feature fusion in contrastive self-supervised learning frameworks on the AUC-ROC for anomaly detection in graphs with missing node attributes. Attribute graphs are…

Abstract: This report synthesises findings from 11 peer-reviewed papers addressing the following research question: How does the performance of self-supervised contrastive learning methods for graph anomaly detection compare to supervised methods in terms of inference efficiency on large-scale graphs with. Combining Graph…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: How does the performance of spectral-based graph anomaly detection methods compare to spatial GNN baselines in terms of F1-score on heterophilic graphs when 50\% of node features are masked. Anomaly detection is…

Abstract: This report synthesises findings from 11 peer-reviewed papers addressing the following research question: How do different graph embedding techniques (e.g., graph attention networks vs. graph convolutional networks) within the Mul-GAD framework impact its performance on cross-domain graph anomaly. Anomaly detection…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: What is the computational efficiency and scalability of Mul-GAD compared to traditional shallow learning methods and other GNN-based approaches when evaluated on large-scale graph anomaly detection. Anomaly…

Abstract: This report synthesises findings from 15 peer-reviewed papers addressing the following research question: How does the semi-supervised approach in Mul-GAD affect its robustness to adversarial attacks compared to fully unsupervised GNN-based anomaly detection methods on heterophilic graphs, measured by. Graph anomaly…

Abstract: This report synthesises findings from 5 peer-reviewed papers addressing the following research question: How does the inference throughput of multimodal segmentation models scale with input resolution on GPU accelerators compared to standard CNN backbones. Multimodal referring segmentation aims to segment target…

Abstract: This report synthesises findings from 4 peer-reviewed papers addressing the following research question: Does Directional Preference Alignment improve the robustness of Code LLMs against syntax errors in low-resource languages more effectively than traditional RLHF approaches. Recently, ChatGPT, along with DALL-E-2…

Abstract: This report synthesises findings from 8 peer-reviewed papers addressing the following research question: How does heterogeneous device latency in federated learning impact the convergence accuracy of multimodal malware detection models compared to homogeneous setups. This work investigates the possibilities enabled…

Abstract: This report synthesises findings from 8 peer-reviewed papers addressing the following research question: How does varying the client participation rate in federated learning impact the zero-shot accuracy of multimodal vision-language models on the VQA-v2 benchmark. It is well understood that client-master…

Abstract: This report synthesises findings from 15 peer-reviewed papers addressing the following research question: What are the trade-offs between model size and code generation quality (measured by HumanEval pass@k) in federated LLMs when trained with partially participating clients under different data. The recent success…

Abstract: This report synthesises findings from 9 peer-reviewed papers addressing the following research question: How does data heterogeneity in federated learning affect the few-shot code generation performance of LLMs (e.g., CodeLlama) on HumanEval, measured by pass@1 and pass@10 scores under varying degrees. Federated…

Abstract: This report synthesises findings from 12 peer-reviewed papers addressing the following research question: How does varying degrees of non-IID data across federated clients influence the inference efficiency and throughput of multimodal models during distributed training. Federated learning learns from scattered data…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: What is the impact of client data heterogeneity on the reasoning capabilities and alignment scores of federated large language models. Federated Learning (FL) enables decentralized training of machine learning…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: How does non-IID data distribution affect the convergence rate and final accuracy of federated language models on code generation benchmarks. The proliferation of edge devices has brought Federated Learning (FL)…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: How does dynamic meta-layer aggregation compare to traditional federated averaging in Byzantine robustness across different attack types (e.g., label flipping, noise injection, backdoor attacks) as. In this…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: Benchmark archaeology: investigate Babilong score discrepancy for GPT-4 — reported 10.0\%–85.0\% (spread 75.0pp) across 2 papers. Sources: 'BABILong: Testing the Limits of LLMs wit' (10.0\%); 'BABILong:. Large…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: What is the impact of random layer aggregation on model alignment in federated learning, as measured by cross-domain performance (e.g., transfer learning accuracy on ImageNet and CIFAR-10) and. The advent of…

Abstract: This report synthesises findings from 8 peer-reviewed papers addressing the following research question: How does random layer aggregation in federated learning models affect inference latency and memory usage on edge devices as measured by throughput (FPS) and GPU/CPU utilization metrics. Performance evaluation is…