Abhinav Girdhar的动态

Check out the new paper, "NeedleBench: Can LLMs Do Retrieval and Reasoning in 1 Million Context Window?" This research introduces NeedleBench, a framework designed to test the retrieval and reasoning capabilities of large language models (LLMs) across extensive text contexts. The paper provides a detailed analysis of the techniques used, making it a valuable resource for developers and researchers in AI and machine learning. The findings highlight significant challenges and opportunities for improving the long-context capabilities of LLMs. For those interested in the technical aspects and practical applications of long-context LLMs, this paper is essential reading. It offers insights and solutions that contribute meaningfully to advancements in the field. Read the full paper here: [https://lnkd.in/ga3WuRmb] #AI #MachineLearning #LLMs #DataScience #Research #Innovation #LongContext #NeedleBench

?????? Our work "N2F2: Hierarchical Scene Understanding with Nested Neural Feature Fields" has been accepted to #ECCV 2024! This will be my first ECCV ?? -- excited to be in Milan! ?? Paper link: https://lnkd.in/d9q9_9SS #eccv24 #eccv2024 #computervision #nerf #neuralradiancefield #ai #ml #research #paper

Humanity’s Last Exam —a new benchmark— shows how far AI still lags behind human expertise. ?? A new benchmark called “Humanity’s Last Exam” has been introduced to evaluate large language models (LLMs) using 3,000 challenging questions across various subjects, including mathematics. Developed by nearly 1,000 subject-matter experts from over 500 institutions worldwide, this benchmark aims to assess LLMs at the frontier of human knowledge. Notably, current state-of-the-art LLMs demonstrate low accuracy on this benchmark, highlighting a significant gap between their capabilities and expert human performance. Paper: https://lnkd.in/edN3uW9A #AI #LLM #Benchmark #ArtificialIntelligence #HumanExpertise

"Excited to share an incredible breakthrough in AI-powered graph and language modeling! ?? Our research introduces GraphAgent, a cutting-edge framework that seamlessly integrates structured and unstructured data to enable both predictive analytics and text generation. By combining the power of graph neural networks and large language models, GraphAgent opens new doors for intuitive, data-driven insights. ???? Explore more: https://lnkd.in/gvFFPwk8 A huge thank you to Chao Huang and the incredible team behind this groundbreaking work for their guidance and collaboration. Together, we’re pushing the boundaries of AI and data integration! ?? #AI #MachineLearning #DataScience #KnowledgeGraphs #LLM #Innovation" Let me know if you'd like to customize this further!

?? Our paper on short text classification with token-level graphs has been accepted for ECIR 2025! ?? ?? In the paper, Gregor Donabauer and Udo Kruschwitz address the classification of short texts in low-resource scenarios. While recent advances in graph machine learning have shown promise in such settings, existing methods often fall short in capturing contextual and semantic information effectively. Our proposed solution constructs text graphs by only using tokens obtained from pre-trained language models (PLMs) which results in a number of positive implications. ???? For more details and the implementations, find a preprint of our paper on arXiv: https://lnkd.in/eHX-EAcH Looking forward to sharing more insights at ECIR 2025! ???? #ECIR2025 #GraphBasedMethods #TextClassification #MachineLearning #AI #PLM #InformationRetrieval???? #Research #ResearchPaperAccepted #InformationScienceRegensburg #StayInformed

Hello Folks! Our spring issue of the #POSTDOCKet is freshly delivered today. Do read some interesting articles on use of Artificial intelligence (AI) in research and digital health and share with your communities. #Happyreading! ??

Hello!! Please read my article “Challenges in Science communication among Postdocs: A non native English speaker’s Perspective” published in the #POSTDOCket by National Postdoctoral Association. Let me know your thoughts.

This paper is an excellent resource for those interested in understanding how reasoning processes function within large language models. ?? The first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1, have been introduced, showcasing groundbreaking advancements in reasoning capabilities for large language models. ?? DeepSeek-R1-Zero, trained solely via reinforcement learning, reveals powerful reasoning behaviors, while DeepSeek-R1 builds on this with multi-stage training to achieve performance comparable to OpenAI-o1-1217. ?? The best part? Both models and six distilled versions (1.5B to 70B parameters) are now open-sourced to empower further innovation in the field. Check them out and explore the future of reasoning in AI! ?? https://lnkd.in/dvenDqRE https://lnkd.in/dtCcNMPX #AI #LLM #Reasoning #OpenSource #Innovation

?? Breaking the Code: AI Language Insights Ever wondered how artificial intelligence truly "understands" language? Join Core CEO Ravi Ganesan as he unravels the fascinating world of transformer architectures and machine learning, revealing how AI transforms complex word patterns into meaningful communication. Discover the magic behind large language models in this must-watch exploration! Watch our full webinar on demand for an in-depth exploration of using AI in behavioral health—available, here: https://hubs.ly/Q02Z_PqM0 #LargeLanguageModels #AI #BehavioralHealth

I had the pleasure of presenting our paper titled "A Robust Quantile Huber Loss with Interpretable Parameter Adjustment in Distributional Reinforcement Learning" at #ICASSP2024 in Seoul last week. We had the opportunity to engage with a diverse group of researchers and received valuable feedback. We extend our gratitude to everyone for their interest in our work. You can access our full paper on arXiv: https://lnkd.in/g2mtEzt2 Additionally, the code, poster, and more information about the paper are available on its webpage: https://lnkd.in/gFSA24Cj #ICASSP2024 #MachineLearning #ReinforcementLearning #AI