Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: What is the correlation between context window length and hallucination rates in multimodal models when evaluated on adversarial noise injections within the MMNeedle dataset. Large Language Models (LLMs) have…

Abstract: This report synthesises findings from 14 peer-reviewed papers addressing the following research question: Does applying semantic guidance in adversarial training improve cross-dataset robustness for code models when evaluated on MBPP versus CodexGLUE. Predicting the trajectories of surrounding objects is a critical…

Abstract: This report synthesises findings from 4 peer-reviewed papers addressing the following research question: Does combining Sentence-T5 and MPNet embeddings improve cross-domain retrieval accuracy on HotpotQA when models are trained exclusively on TriviaQA. Modern information retrieval (IR) models, trained exclusively on…

Abstract: This report synthesises findings from 4 peer-reviewed papers addressing the following research question: How do graph-enhanced multimodal models scale in terms of throughput versus accuracy compared to flat fusion models under non-adversarial conditions on MM-Vet. Robot vision has greatly benefited from advancements…

Abstract: This report synthesises findings from 1 peer-reviewed paper addressing the following research question: Does the use of hybrid embeddings in Tree of Reviews improve robustness against distractor documents in multi-hop QA datasets compared to single-embedding retrieval methods. The Portable Document Format (PDF) is…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: How does the inference latency of horizon-adaptive VLA policies in LongNav-R1 compare to fixed-horizon baselines across varying navigation episode lengths. This paper develops LongNav-R1, an end-to-end multi-turn…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: What is the impact of horizon-adaptive multi-turn reinforcement learning on the robustness of VLA navigation policies against visual obscurations in simulated 3D environments. This paper develops LongNav-R1, an…

Abstract: This report synthesises findings from 15 peer-reviewed papers addressing the following research question: How does the multi-turn iterative preference learning approach compare to Supervised Fine-Tuning (SFT) in improving zero-shot generalization of LLMs on mathematical reasoning tasks, as measured by. Recent studies…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: How does the sample efficiency of multi-turn RL frameworks like LongNav-R1 compare to single-turn baselines when scaling Vision-Language-Action models on the Habitat-3D benchmark. This paper develops LongNav-R1,…

Abstract: This report synthesises findings from 14 peer-reviewed papers addressing the following research question: How does the sample efficiency of multi-turn RL training for VLA navigation compare to single-turn imitation learning as trajectory length increases. This paper develops LongNav-R1, an end-to-end multi-turn…

Abstract: This report synthesises findings from 9 peer-reviewed papers addressing the following research question: How does the inference efficiency of LongNav-R1's horizon-adaptive framework scale with increasing environmental complexity in Habitat 2.0 compared to other multi-turn RL methods, measured in frames. This paper…

Abstract: This report synthesises findings from 4 peer-reviewed papers addressing the following research question: What is the impact of dynamically scaling the number of generated unit tests per candidate solution on the training stability and final pass@k scores of reinforcement learning from execution feedback. Optimistic…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: What is the efficiency trade-off in terms of training throughput and sample complexity when comparing DPO with rationales versus standard DPO on the LLaVA-Bench benchmark, evaluated by the number of. Aligning…

Abstract: This report synthesises findings from 4 peer-reviewed papers addressing the following research question: Does integrating explicit rationales into preference data reduce the variance in pass@1 scores for hard-tier GSM8K problems compared to standard DPO. Aligning language models with human preferences through…

Abstract: This report synthesises findings from 15 peer-reviewed papers addressing the following research question: Does scaling the number of preference data samples in difficulty-based DPO improve robustness to adversarial perturbations in self-invoking code generation, as evaluated on a modified HumanEval Pro. Aligning…

Abstract: This report synthesises findings from 3 peer-reviewed papers addressing the following research question: What is the impact of sliding window attention mechanisms on inference throughput and memory usage for sequence lengths exceeding 32k tokens in code LLMs. Recent advances in language modeling have demonstrated the…

Abstract: This report synthesises findings from 8 peer-reviewed papers addressing the following research question: To what extent do synthetic graph generation techniques improve the generalization of graph neural networks in low-density regimes compared to standard train-test splits. Graph Neural Networks (GNNs) are one of…

Abstract: This report synthesises findings from 7 peer-reviewed papers addressing the following research question: How does the choice of data augmentation strategy impact the robustness of F1 and AUC metrics for graph anomaly detection models across varying graph densities. Detecting anomalies in data is a vital task, with…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: How does chain-of-thought prompting impact cross-domain reasoning accuracy of LLMs when transferring from financial text corpora to legal document analysis. Financial news sentiment analysis is crucial for…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: What is the inference latency overhead of multimodal alignment techniques in large vision-language models when evaluated on low-resource hardware benchmarks. Robot vision has greatly benefited from advancements…

Abstract: This report synthesises findings from 13 peer-reviewed papers addressing the following research question: To what extent does code generation performance degrade in LLMs when fine-tuned on proprietary software repositories versus open-source GitHub datasets. Large Language Models (LLMs) have demonstrated significant…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: How does the sample efficiency of LightGCL compare to SGL and GCA in low-data regimes on MovieLens and Amazon recommendation benchmarks. Graph neural network (GNN) is a powerful learning approach for graph-based…

Abstract: This report synthesises findings from 16 peer-reviewed papers addressing the following research question: How does the efficiency and accuracy of graph contrastive learning methods scale with increasing sparsity in interaction graphs, as evaluated by HR and MAP metrics on large-scale datasets. Contrastive learning…

Abstract: This report synthesises findings from 10 peer-reviewed papers addressing the following research question: Do simplified augmentation pipelines in graph contrastive learning maintain robust performance across different sparsity levels in interaction graphs when measured by HR and MAP compared to complex. Attributed…

Abstract: This report synthesises findings from 7 peer-reviewed papers addressing the following research question: How does the adversarial robustness of XSimGCL compare to other graph contrastive learning methods when evaluated on the Yelp and Amazon review datasets using NDCG@10 and NDCG@20 as metrics. Contrastive learning…