Abstract: This report synthesises findings from 16 peer-reviewed papers addressing the following research question: How does varying the margin parameter in contrastive learning objectives impact retrieval accuracy in cross-domain dense retrievers on NQ compared to models trained solely with pairwise ranking losses. 9 claims…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: How does manifold-aware adversarial training impact the zero-shot image classification accuracy of CLIP on ImageNet-1K compared to standard adversarial fine-tuning. 10 claims were extracted from source…

Abstract: This report synthesises findings from 14 peer-reviewed papers addressing the following research question: Do dual-encoder models exhibit greater latency-efficiency tradeoffs than cross-encoders when evaluated for robustness against adversarial token perturbations in open-domain QA tasks. 12 claims were extracted from…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: What is the impact of synthetic misspelling augmentation on the retrieval accuracy of dual-encoder versus cross-encoder architectures in the MTEB evaluation framework. 7 claims were extracted from source…

Abstract: This report synthesises findings from 8 peer-reviewed papers addressing the following research question: How does character-level noise robustness in dual-encoder retrieval models compare to cross-encoders on the BEIR benchmark under varying typo severities. 9 claims were extracted from source literature; 9 were…

Abstract: This report synthesises findings from 14 peer-reviewed papers addressing the following research question: Does incorporating synthetic misspelling augmentation during contrastive learning improve cross-lingual retrieval performance while maintaining acceptable throughput on large-scale language model. 11 claims were…

Abstract: This report synthesises findings from 9 peer-reviewed papers addressing the following research question: What is the impact of typo-injected adversarial training on the robustness of multimodal retrieval systems when evaluated against noisy query datasets. 8 claims were extracted from source literature; 8 were…

Abstract: This report synthesises findings from 14 peer-reviewed papers addressing the following research question: How does adversarial training with synthetic misspellings affect the retrieval accuracy and inference latency of contrastive learning models on standard text retrieval benchmarks. 14 claims were extracted from…

Abstract: This report synthesises findings from 12 peer-reviewed papers addressing the following research question: Does the defense strategy utilizing one-to-many image-text mappings improve cross-domain generalization robustness against multimodal adversarial perturbations in zero-shot retrieval tasks. 9 claims were…

Abstract: This report synthesises findings from 11 peer-reviewed papers addressing the following research question: What is the impact of one-to-many relationship augmentation on the inference latency and throughput of vision-language models during adversarial training on large-scale datasets. 9 claims were extracted from…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: How does leveraging one-to-many image-text relationships affect the robustness accuracy of CLIP-based models under gradient-based multimodal adversarial attacks compared to standard contrastive loss. 11 claims…

Abstract: This report synthesises findings from 4 peer-reviewed papers addressing the following research question: How do geodesic distance-based dense retrievers perform compared to Euclidean-based models in terms of robustness to adversarial perturbations or noisy queries in the BEIR-NL benchmark. 7 claims were extracted…

Abstract: This report synthesises findings from 1 peer-reviewed paper addressing the following research question: How does the cross-lingual transferability of geodesic distance-based dense retrievers compare to Euclidean-based models when evaluated on multilingual BEIR benchmarks. 6 claims were extracted from source…

Abstract: This report synthesises findings from 9 peer-reviewed papers addressing the following research question: What is the impact of model size scaling on the inference efficiency and retrieval accuracy of geodesic distance-based dense retrievers versus Euclidean-based models in Dutch IR tasks. 7 claims were extracted from…

Abstract: This report synthesises findings from 11 peer-reviewed papers addressing the following research question: What is the impact of domain shift on the zero-shot retrieval performance of dense retrievers when trained with geodesic distance versus Euclidean distance on the BEIR benchmark. 13 claims were extracted from…

Abstract: This report synthesises findings from 9 peer-reviewed papers addressing the following research question: How does the computational efficiency of dense retrievers using geodesic distance compare to Euclidean distance when scaling to large-scale datasets in language model benchmarks. 10 claims were extracted from…

Abstract: This report synthesises findings from 5 peer-reviewed papers addressing the following research question: What is the impact of different hyperbolic space embeddings (e.g., Poincar, Lorentz) on the retrieval efficiency and throughput of zero-shot cross-lingual models in XOR-TyDi QA, measured by recall@k. 0 claims were…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: Does incorporating geodesic distance metrics into cross-encoder reranking improve retrieval accuracy on the BEIR benchmark compared to traditional Euclidean-based methods. 9 claims were extracted from source…

Abstract: This report synthesises findings from 15 peer-reviewed papers addressing the following research question: What is the impact of different contrastive loss functions on the robustness of hyperbolic vs. Euclidean embeddings for cross-lingual retrieval in XOR-TyDi QA when evaluated under noisy or. 10 claims were…

Abstract: This report synthesises findings from 14 peer-reviewed papers addressing the following research question: Can multitask learning strategies derived from radar HRRP recognition enhance the robustness of multimodal models against adversarial perturbations in low-resource code generation benchmarks. 9 claims were…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: How does multitask learning with shared aspect-dependent parameters improve few-shot cross-lingual retrieval accuracy on XOR-TyDi QA compared to single-task hyperbolic contrastive models. 0 claims were extracted…

Abstract: This report synthesises findings from 12 peer-reviewed papers addressing the following research question: Does replacing cosine similarity with geodesic distance in dense retrievers improve adversarial robustness on BEIR out-of-distribution subsets compared to standard contrastive learning baselines. 0 claims were…

Abstract: This report synthesises findings from 11 peer-reviewed papers addressing the following research question: How does the performance of NMIXX compare to domain-adaptive cross-lingual embeddings like LaBSE or XLM-R on financial QA benchmarks in Korean when evaluated using semantic similarity metrics beyond. 11 claims…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: To what extent do multimodal embeddings improve the robustness of manifold-aware distance metrics in dense retrieval when evaluated on cross-modal MTEB tasks involving text, images, and structured. 0 claims were…

Abstract: This report synthesises findings from 12 peer-reviewed papers addressing the following research question: What is the impact of increasing video demonstration duration within the context window on the reasoning error rate of Gemini 1.5 models in multimodal code synthesis benchmarks. 5 claims were extracted from…