Abstract: This report synthesises findings from 15 peer-reviewed papers addressing the following research question: What is the impact of evidential uncertainty quantification on the robustness of multimodal models like CLIP against domain shift and adversarial perturbations in zero-shot classification tasks. 6 claims were…

Abstract: This report synthesises findings from 14 peer-reviewed papers addressing the following research question: What is the impact of multimodal fusion strategies on inference latency and throughput for real-time Visual Question Answering systems on edge devices. 9 claims were extracted from source literature; 8 were…

Abstract: This report synthesises findings from 9 peer-reviewed papers addressing the following research question: How do manifold-aware distance metrics perform in cross-modal retrieval tasks (e.g., text-to-image) compared to cosine similarity when using decoder-only LLMs on multimodal benchmarks like MSCOCO or. 6 claims were…

Abstract: This report synthesises findings from 14 peer-reviewed papers addressing the following research question: How does evidential deep learning with non-negative evidence constraints affect cross-modal retrieval accuracy on CLIP and ALBEF compared to standard softmax-based metric learning. 5 claims were extracted from…

Abstract: This report synthesises findings from 3 peer-reviewed papers addressing the following research question: What is the effect of scaling the dimensionality of embeddings on the retrieval accuracy gap between manifold-aware distance metrics and cosine similarity in out-of-domain BEIR tasks. 6 claims were extracted from…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: Can cross-encoder models trained with manifold-aware objectives outperform traditional dense retrievers on adversarial benchmarks like Adversarial NQ while maintaining competitive accuracy on. 10 claims were…

Abstract: This report synthesises findings from 1 peer-reviewed paper addressing the following research question: How does cross-domain fine-tuning affect the robustness of dense retrieval models to misspellings and noise, as evaluated on domain-specific benchmarks like BEIR or FEVER, measuring changes in. 6 claims were…

Abstract: This report synthesises findings from 5 peer-reviewed papers addressing the following research question: How does the efficiency of manifold-aware fine-tuning techniques compare to Euclidean-based models when evaluated on large-scale adversarial query benchmarks like ANLIR in terms of inference latency. 7 claims were…

Abstract: This report synthesises findings from 15 peer-reviewed papers addressing the following research question: Does integrating manifold-aware loss functions improve the zero-shot cross-domain generalization of dual-encoder retrievers on diverse NLU benchmarks without compromising training convergence speed. 10 claims…

Abstract: This report synthesises findings from 14 peer-reviewed papers addressing the following research question: What is the impact of manifold regularization on the robustness of dense retrieval models against adversarial token perturbations compared to standard dual-encoder architectures. 9 claims were extracted from…

Abstract: This report synthesises findings from 14 peer-reviewed papers addressing the following research question: To what extent do manifold-aware embeddings improve robustness against adversarial query perturbations in cross-domain retrieval tasks relative to domain-adaptive models on the BEIR benchmark. 9 claims were…

Abstract: This report synthesises findings from 7 peer-reviewed papers addressing the following research question: How does the cross-domain retrieval accuracy of manifold-aware embedding projections compare to domain-adaptive fine-tuning on the BEIR benchmark when evaluated on low-resource languages. 8 claims were extracted…

Abstract: This report synthesises findings from 12 peer-reviewed papers addressing the following research question: What is the impact of manifold-aware projection layers on inference latency and throughput compared to full parameter fine-tuning in large-scale dense retrieval systems evaluated on the MS MARCO. 5 claims were…

Abstract: This report synthesises findings from 15 peer-reviewed papers addressing the following research question: What is the throughput degradation when scaling nonlinear dimensionality reduction techniques for Bayesian optimization in large-scale multimodal embedding spaces. 9 claims were extracted from source literature;…

Abstract: This report synthesises findings from 16 peer-reviewed papers addressing the following research question: How does the inference latency of VAE-based latent space optimization compare to random projection methods in high-dimensional multimodal recommendation tasks on LAION-5B. 9 claims were extracted from source…

Abstract: This report synthesises findings from 9 peer-reviewed papers addressing the following research question: How does the computational efficiency of manifold-regularized dense retrievers compare to standard dual-encoder models during inference on large-scale benchmarks like BEIR. 11 claims were extracted from source…

Abstract: This report synthesises findings from 15 peer-reviewed papers addressing the following research question: What is the effect of manifold-aware projection techniques on the robustness of retrieval-augmented generation systems when evaluated on out-of-domain technical regulation datasets. 10 claims were extracted from…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: How does domain-adaptive fine-tuning of large language models impact MRR@10 scores on legal text retrieval benchmarks compared to general-purpose pre-trained models. 10 claims were extracted from source…

Abstract: This report synthesises findings from 15 peer-reviewed papers addressing the following research question: What is the effect of varying the size of the training corpus on the OOD generalization capabilities of dense retrievers when evaluated on BEIR's heterogeneous datasets. 7 claims were extracted from source…

Abstract: This report synthesises findings from 16 peer-reviewed papers addressing the following research question: How does the integration of contrastive learning objectives in dense retrievers impact zero-shot performance on cross-domain tasks like NQ compared to models using only standard pairwise ranking. 7 claims were…

Abstract: This report synthesises findings from 11 peer-reviewed papers addressing the following research question: How do different architectures (e.g., transformer vs. hybrid models) affect the trade-off between robustness to misspellings and inference efficiency in dense retrieval systems. 10 claims were extracted from…

Abstract: This report synthesises findings from 4 peer-reviewed papers addressing the following research question: Does integrating manifold-aware defense mechanisms into vision-language models degrade inference throughput or zero-shot classification performance on standard benchmarks like ImageNet and MSCOCO. 11 claims were…

Abstract: This report synthesises findings from 6 peer-reviewed papers addressing the following research question: How does the robustness of dual-encoder retrieval models to misspellings compare to cross-encoder models, as measured by HELM fairness and robustness metrics across different misspelling severities. 11 claims were…

Abstract: This report synthesises findings from 11 peer-reviewed papers addressing the following research question: What is the impact of incorporating adversarial training with synthetic misspellings on the accuracy and inference latency of contrastive learning-based retrieval systems. 6 claims were extracted from source…

Abstract: This report synthesises findings from 11 peer-reviewed papers addressing the following research question: What is the impact of leveraging one-to-many image-text relationships on the retrieval accuracy of vision-language models when evaluated against gradient-based multimodal adversarial attacks. 0 claims were…