✨VideoAtlas: Navigating Long-Form Video in Logarithmic Compute

📝 Summary:
VideoAtlas enables lossless video representation through hierarchical grids, enabling efficient long-context processing via recursive language models with adaptive compute allocation. AI-generated sum...

🔹 Publication Date: Published on Mar 18

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.17948
• PDF: https://arxiv.org/pdf/2603.17948
• Project Page: https://mohammad2012191.github.io/VideoAtlas/
• Github: https://github.com/mohammad2012191/VideoAtlas

==================================

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Efficient Training-Free Multi-Token Prediction via Embedding-Space Probing

📝 Summary:
A training-free method for multi-token prediction in large language models using mask tokens from the embedding space enables parallel token generation with improved throughput and accuracy. AI-genera...

🔹 Publication Date: Published on Mar 18

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.17942
• PDF: https://arxiv.org/pdf/2603.17942

==================================

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨AI Scientist via Synthetic Task Scaling

📝 Summary:
This paper introduces a novel synthetic environment pipeline that generates machine learning challenges for training AI agents. Student models trained with these synthetic tasks, using teacher trajectories, achieve significantly improved performance on MLGym benchmarks.

🔹 Publication Date: Published on Mar 17

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.17216
• PDF: https://arxiv.org/pdf/2603.17216

==================================

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AIResearch #MachineLearning #SyntheticTasks #AIAgents #DeepLearning

✨GigaWorld-Policy: An Efficient Action-Centered World--Action Model

📝 Summary:
GigaWorld-Policy is an action-centered World-Action Model that significantly improves robotic policy learning. It decouples visual and motion representations, using dual supervision from action prediction and video generation. This allows for 9x faster inference and 7% higher task success rates c...

🔹 Publication Date: Published on Mar 18

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.17240
• PDF: https://arxiv.org/pdf/2603.17240
• Project Page: https://gigaai-research.github.io/GigaWorld-Policy/
• Github: https://github.com/open-gigaai/giga-world-policy

==================================

For more data science resources:
✓ https://t.iss.one/DataScienceT

#Robotics #MachineLearning #WorldModels #DeepLearning #PolicyLearning

✨Fanar-Sadiq: A Multi-Agent Architecture for Grounded Islamic QA

📝 Summary:
Fanar-Sadiq is a bilingual multi-agent Islamic assistant addressing LLM inaccuracies in religious QA. It uses a tool-using architecture with specialized modules for diverse queries like scripture, fiqh, and calculations, ensuring grounded, accurate, and deterministic answers.

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.08501
• PDF: https://arxiv.org/pdf/2603.08501

==================================

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨PPTAgent: Generating and Evaluating Presentations Beyond Text-to-Slides

📝 Summary:
PPTAgent, a two-stage approach, improves presentation generation by analyzing reference presentations and ensuring structural and content consistency, outperforming traditional methods across content,...

🔹 Publication Date: Published on Jan 7, 2025

🔹 Paper Links:
• arXiv Page: https://huggingface.co/collections/ICIP/pptagent
• PDF: https://arxiv.org/pdf/2501.03936
• Project Page: https://github.com/icip-cas/PPTAgent
• Github: https://github.com/icip-cas/PPTAgent

✨ Datasets citing this paper:
• https://huggingface.co/datasets/Forceless/Zenodo10K

==================================

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Expert Threshold Routing for Autoregressive Language Modeling with Dynamic Computation Allocation and Load Balancing

📝 Summary:
Expert Threshold ET routing dynamically allocates computation in MoE models. Tokens route to experts based on individual scores exceeding EMA thresholds, achieving load balance without auxiliary losses. ET lowers cross-entropy loss by 0.067 compared to Token-choice MoE.

🔹 Publication Date: Published on Mar 12

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.11535
• PDF: https://arxiv.org/pdf/2603.11535
• Github: https://github.com/MasterGodzilla/Expert-Threshold-Routing

==================================

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨From Prior to Pro: Efficient Skill Mastery via Distribution Contractive RL Finetuning

📝 Summary:
DICE-RL refines pretrained generative robot policies via reinforcement learning distribution contraction. It boosts high-success behaviors, leading to stable, sample-efficient mastery of complex manipulation from pixels on real robots.

🔹 Publication Date: Published on Mar 10

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.10263
• PDF: https://arxiv.org/pdf/2603.10263
• Project Page: https://zhanyisun.github.io/dice.rl.2026/

==================================

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨AdapterTune: Zero-Initialized Low-Rank Adapters for Frozen Vision Transformers

📝 Summary:
AdapterTune introduces zero-initialized low-rank adapters for Vision Transformers, addressing optimization instability and capacity issues. This method prevents representation drift and significantly improves accuracy, often outperforming full fine-tuning with fewer parameters.

🔹 Publication Date: Published on Mar 16

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.14706
• PDF: https://arxiv.org/pdf/2603.14706
• Github: https://github.com/salimkhazem/adaptertune

==================================

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨OSM-based Domain Adaptation for Remote Sensing VLMs

📝 Summary:
A self-contained domain adaptation framework for vision-language models in remote sensing uses OpenStreetMap data and optical character recognition to generate captions without requiring external teac...

🔹 Publication Date: Published on Mar 12

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.11804
• PDF: https://arxiv.org/pdf/2603.11804

==================================

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research