A Caltech researcher presented at MBZUAI on memory representation and retrieval, contrasting AI and neuroscience approaches. Current AI retrieval systems like RAG retrieve via fine-tuning and embedding similarity, while the presenter argued for exploring retrieval via combinatorial object identity or spatial proximity. The research explores circuit-level retrieval via domain fine-tuned LLMs and distributed memory for image retrieval using semantic similarity. Why it matters: The work suggests structured databases and retrieval-focused training can allow smaller models to outperform larger general-purpose models, offering efficiency gains for AI development in the region.
Caltech graduate student Surya Narayanan Hari presented his research on replicating human-like memory in machines at MBZUAI. He discussed how the thalamus, which filters sensory and motor signals in the brain, inspires the development of routed monolithic models in AI. Hari explained that memory retrieval occurs on object, embedding, and circuit levels in the human brain. Why it matters: This talk highlights the potential of neuroscience-inspired AI architectures for improving memory and information processing in AI systems, which could accelerate the development of more efficient and context-aware AI models in the region.
This article discusses retrieval augmentation in text generation, where information retrieved from an external source is used to condition predictions. It references recent work on retrieval-augmented image captioning, showing that model size can be greatly reduced when training data is available through retrieval. The author intends to continue this work focusing on the intersection of retrieval augmentation and in-context learning, and controllable image captioning for language learning materials. Why it matters: This research direction has the potential to improve transfer learning in vision-language models, which could be especially relevant for downstream applications in Arabic NLP and multimodal tasks.
Prof. Chun Jason Xue from the City University of Hong Kong presented research on optimizing mobile memory and storage by analyzing mobile application characteristics, noting their differences from server applications. The research explores system software designs inherited from the Linux kernel and identifies optimization opportunities in mobile memory and storage management. Xue's work aims to enhance user experience on mobile devices through mobile application characterization, focusing on non-volatile and flash memories. Why it matters: Optimizing mobile systems based on the unique characteristics of mobile applications can significantly improve device performance and user experience in the region.
A novel agent-based framework called FIRE is introduced for fact-checking long-form text. FIRE iteratively integrates evidence retrieval and claim verification, deciding whether to provide a final answer or generate a subsequent search query. Experiments show FIRE achieves comparable performance to existing methods while reducing LLM costs by 7.6x and search costs by 16.5x.
A new paper at ICCV 2025, co-authored by MBZUAI Ph.D. student Dmitry Demidov, introduces Dense-WebVid-CoVR, a 1.6-million sample benchmark for composed video retrieval (CoVR). The benchmark features longer, context-rich descriptions and modification texts, generated using Gemini Pro and GPT-4o, with manual verification. The paper also presents a unified fusion approach that jointly reasons across video and text inputs, improving performance on fine-grained edit details. Why it matters: This work advances video search capabilities by enabling more human-like queries, which is crucial for creative and analytic workflows that require nuanced video retrieval.
This paper introduces MOTOR, a multimodal retrieval and re-ranking approach for medical visual question answering (MedVQA) that uses grounded captions and optimal transport to capture relationships between queries and retrieved context, leveraging both textual and visual information. MOTOR identifies clinically relevant contexts to augment VLM input, achieving higher accuracy on MedVQA datasets. Empirical analysis shows MOTOR outperforms state-of-the-art methods by an average of 6.45%.
A CMU professor and MBZUAI affiliated faculty presented research on how LLMs store and use knowledge learned during pre-training. The study used a synthetic biography dataset to show that LLMs may not effectively use memorized knowledge at inference time, even with zero training loss. Data augmentation during pre-training can force the model to store knowledge in specific token embeddings. Why it matters: The research highlights limitations in LLM knowledge manipulation and extraction, with implications for improving model architectures and training strategies for more effective knowledge utilization in Arabic LLMs.