✨How Much 3D Do Video Foundation Models Encode?

📝 Summary:
A new framework quantifies 3D understanding in Video Foundation Models VidFMs. VidFMs, trained only on video, show strong 3D awareness, often surpassing expert 3D models, providing insights for 3D AI.

🔹 Publication Date: Published on Dec 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.19949
• PDF: https://arxiv.org/pdf/2512.19949
• Project Page: https://vidfm-3d-probe.github.io/
• Github: https://vidfm-3d-probe.github.io

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

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

#VideoFoundationModels #3DUnderstanding #ComputerVision #AIResearch #DeepLearning

✨Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass

📝 Summary:
Fast3R is a Transformer-based method for efficient and scalable multi-view 3D reconstruction. It processes many images in parallel in a single forward pass, improving speed and accuracy over pairwise approaches like DUSt3R.

🔹 Publication Date: Published on Jan 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2501.13928
• PDF: https://arxiv.org/pdf/2501.13928
• Github: https://github.com/naver/dust3r/pull/16

🔹 Models citing this paper:
• https://huggingface.co/jedyang97/Fast3R_ViT_Large_512

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

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

#3DReconstruction #ComputerVision #Transformers #Fast3R #DeepLearning

✨InsertAnywhere: Bridging 4D Scene Geometry and Diffusion Models for Realistic Video Object Insertion

📝 Summary:
InsertAnywhere is a framework for realistic video object insertion. It uses 4D aware mask generation for geometric consistency and an extended diffusion model for appearance-faithful synthesis, outperforming existing methods.

🔹 Publication Date: Published on Dec 19

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.17504
• PDF: https://arxiv.org/pdf/2512.17504
• Project Page: https://myyzzzoooo.github.io/InsertAnywhere/
• Github: https://github.com/myyzzzoooo/InsertAnywhere

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

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

#VideoEditing #DiffusionModels #ComputerVision #DeepLearning #GenerativeAI

✨See Less, See Right: Bi-directional Perceptual Shaping For Multimodal Reasoning

📝 Summary:
Bi-directional Perceptual Shaping BiPS improves vision-language models by using question-conditioned masked views to shape perception during training. It employs two constraints to ensure complete coverage of relevant pixels and enforce fine-grained visual reliance, preventing text-only shortcuts...

🔹 Publication Date: Published on Dec 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.22120
• PDF: https://arxiv.org/pdf/2512.22120
• Github: https://github.com/zss02/BiPS

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

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

#MultimodalAI #VisionLanguageModels #MachineLearning #AIResearch #DeepLearning

✨TimeBill: Time-Budgeted Inference for Large Language Models

📝 Summary:
TimeBill is a framework for LLMs in time-critical systems. It predicts execution time and adaptively adjusts KV cache eviction to balance inference efficiency and response performance within given time budgets, improving task completion rates.

🔹 Publication Date: Published on Dec 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.21859
• PDF: https://arxiv.org/pdf/2512.21859

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

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

#LLM #AI #RealTimeAI #InferenceOptimization #DeepLearning

✨UniPercept: Towards Unified Perceptual-Level Image Understanding across Aesthetics, Quality, Structure, and Texture

📝 Summary:
UniPercept-Bench provides a unified framework and datasets for perceptual image understanding aesthetics, quality, structure, texture. The UniPercept model, trained with DAPT and T-ARL, outperforms MLLMs, generalizes across VR and VQA, and acts as a text-to-image reward model.

🔹 Publication Date: Published on Dec 25

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.21675
• PDF: https://arxiv.org/pdf/2512.21675
• Project Page: https://thunderbolt215.github.io/Unipercept-project/
• Github: https://github.com/thunderbolt215/UniPercept

🔹 Models citing this paper:
• https://huggingface.co/Thunderbolt215215/UniPercept

✨ Datasets citing this paper:
• https://huggingface.co/datasets/Thunderbolt215215/UniPercept-Bench

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

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

#ImageUnderstanding #ComputerVision #AIResearch #PerceptualAI #DeepLearning

✨Omni-Weather: Unified Multimodal Foundation Model for Weather Generation and Understanding

📝 Summary:
Omni-Weather is a new multimodal foundation model that unifies weather generation and understanding in a single architecture. It uses shared self-attention and a Chain-of-Thought dataset for interpretable, high-quality outputs, achieving state-of-the-art performance.

🔹 Publication Date: Published on Dec 25

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.21643
• PDF: https://arxiv.org/pdf/2512.21643

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

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

#WeatherGeneration #FoundationModels #MultimodalAI #AIResearch #DeepLearning

✨Coupling Experts and Routers in Mixture-of-Experts via an Auxiliary Loss

📝 Summary:
Expert-Router Coupling ERC loss aligns MoE router decisions with expert capabilities. It uses proxy tokens and activation constraints to ensure experts specialize, improving performance and computational efficiency. ERC also allows tracking expert specialization during training.

🔹 Publication Date: Published on Dec 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.23447
• PDF: https://arxiv.org/pdf/2512.23447

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

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

#MixtureOfExperts #DeepLearning #MachineLearning #AI #NeuralNetworks

✨Yume-1.5: A Text-Controlled Interactive World Generation Model

📝 Summary:
Yume-1.5 is a novel framework that generates realistic, interactive, and continuous worlds from a single image or text prompt. It overcomes prior limitations in real-time performance and text control by using unified context compression, streaming acceleration, and text-controlled world events.

🔹 Publication Date: Published on Dec 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.22096
• PDF: https://arxiv.org/pdf/2512.22096
• Project Page: https://stdstu12.github.io/YUME-Project/
• Github: https://github.com/stdstu12/YUME

🔹 Models citing this paper:
• https://huggingface.co/stdstu123/Yume-5B-720P

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

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

#AI #GenerativeAI #WorldGeneration #ComputerGraphics #DeepLearning

✨SpotEdit: Selective Region Editing in Diffusion Transformers

📝 Summary:
SpotEdit is a training-free framework for selective image editing in diffusion transformers. It avoids reprocessing stable regions by reusing their features, combining them with edited areas. This reduces computation and preserves unchanged regions, enhancing efficiency and precision.

🔹 Publication Date: Published on Dec 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.22323
• PDF: https://arxiv.org/pdf/2512.22323
• Project Page: https://biangbiang0321.github.io/SpotEdit.github.io
• Github: https://biangbiang0321.github.io/SpotEdit.github.io

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

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

#ImageEditing #DiffusionModels #ComputerVision #AIResearch #DeepLearning

✨Evaluating Parameter Efficient Methods for RLVR

📝 Summary:
This work evaluates 12 PEFT methods for RLVR in mathematical reasoning, challenging LoRAs default use. It finds that structural variants like DoRA outperform LoRA, while SVD-informed methods fail and extreme parameter reduction bottlenecks reasoning.

🔹 Publication Date: Published on Dec 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.23165
• PDF: https://arxiv.org/pdf/2512.23165

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

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

#PEFT #RLVR #MathematicalReasoning #LoRA #DeepLearning

✨UltraShape 1.0: High-Fidelity 3D Shape Generation via Scalable Geometric Refinement

📝 Summary:
UltraShape 1.0 is a 3D diffusion framework that generates high-fidelity shapes using a two-stage process: coarse then refined geometry. It includes a novel data pipeline improving dataset quality, enabling strong geometric results on public data.

🔹 Publication Date: Published on Dec 24

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.21185
• PDF: https://arxiv.org/pdf/2512.21185
• Project Page: https://pku-yuangroup.github.io/UltraShape-1.0/
• Github: https://pku-yuangroup.github.io/UltraShape-1.0/

🔹 Models citing this paper:
• https://huggingface.co/infinith/UltraShape

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

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

#3DGeneration #DiffusionModels #GenerativeAI #ComputerGraphics #DeepLearning

✨CosineGate: Semantic Dynamic Routing via Cosine Incompatibility in Residual Networks

📝 Summary:
CosineGate enables dynamic routing in residual networks using cosine incompatibility to skip redundant blocks. This reduces computation by up to 28.5 percent while matching or exceeding ResNet-20 accuracy, without auxiliary supervision.

🔹 Publication Date: Published on Dec 21, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.22206
• PDF: https://arxiv.org/pdf/2512.22206
• Github: https://github.com/thotayogeswarreddy/CosineGate

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

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

#DeepLearning #NeuralNetworks #DynamicRouting #ModelEfficiency #AIResearch

✨Youtu-LLM: Unlocking the Native Agentic Potential for Lightweight Large Language Models

📝 Summary:
Youtu-LLM is a lightweight 1.96B LLM, pre-trained from scratch with a compact architecture and a multi-stage curriculum focused on commonsense, STEM, and agentic tasks. It achieves state-of-the-art performance for sub-2B models, demonstrating strong intrinsic agentic capabilities.

🔹 Publication Date: Published on Dec 31, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.24618
• PDF: https://arxiv.org/pdf/2512.24618

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

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

#LLM #AI #AgenticAI #LightweightLLM #DeepLearning

✨SpaceTimePilot: Generative Rendering of Dynamic Scenes Across Space and Time

📝 Summary:
SpaceTimePilot is a video diffusion model for dynamic scene rendering, offering independent control over spatial viewpoint and temporal motion. It achieves precise space-time disentanglement via a time-embedding, temporal-warping training, and a synthetic dataset.

🔹 Publication Date: Published on Dec 31, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.25075
• PDF: https://arxiv.org/pdf/2512.25075
• Project Page: https://zheninghuang.github.io/Space-Time-Pilot/

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

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

#VideoDiffusion #GenerativeAI #DynamicScenes #ComputerGraphics #DeepLearning

✨Geometry-Aware Optimization for Respiratory Sound Classification: Enhancing Sensitivity with SAM-Optimized Audio Spectrogram Transformers

📝 Summary:
This paper improves respiratory sound classification using AST enhanced with SAM. It optimizes loss surface geometry for flatter minima, yielding state-of-the-art 68.10% score and crucial 68.31% sensitivity on ICBHI 2017.

🔹 Publication Date: Published on Dec 27, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.22564
• PDF: https://arxiv.org/pdf/2512.22564

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

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

#RespiratoryHealth #MedicalAI #DeepLearning #SoundClassification #AIHealthcare

✨mHC: Manifold-Constrained Hyper-Connections

📝 Summary:
Manifold-Constrained Hyper-Connections mHC resolve training instability and scalability issues of Hyper-Connections HC. mHC restores identity mapping via manifold projection and infrastructure optimization, enabling effective large-scale training with improved performance.

🔹 Publication Date: Published on Dec 31, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.24880
• PDF: https://arxiv.org/pdf/2512.24880

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

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

#MachineLearning #DeepLearning #NeuralNetworks #ManifoldLearning #AI

✨Kronos: A Foundation Model for the Language of Financial Markets

📝 Summary:
Kronos is a novel foundation model for financial K-line data. It uses a specialized tokenizer and autoregressive pre-training on a vast dataset to significantly outperform existing models in price and volatility forecasting, and synthetic data generation, establishing it as a versatile tool for f...

🔹 Publication Date: Published on Aug 2, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2508.02739
• PDF: https://arxiv.org/pdf/2508.02739
• Github: https://github.com/shiyu-coder/Kronos

🔹 Models citing this paper:
• https://huggingface.co/NeoQuasar/Kronos-base
• https://huggingface.co/NeoQuasar/Kronos-Tokenizer-base
• https://huggingface.co/NeoQuasar/Kronos-mini

✨ Spaces citing this paper:
• https://huggingface.co/spaces/ByronWang2005/Kronos-CS2-Skins-Forecast-Demo
• https://huggingface.co/spaces/yangyang158/kronos
• https://huggingface.co/spaces/heyunfei/crypt

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

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

#FoundationModel #FinancialAI #DeepLearning #QuantitativeFinance #Forecasting

✨Guiding a Diffusion Transformer with the Internal Dynamics of Itself

📝 Summary:
This paper introduces Internal Guidance IG for diffusion models, which adds auxiliary supervision to intermediate layers during training and extrapolates outputs during sampling. This simple strategy significantly improves training efficiency and generation quality. IG achieves state-of-the-art F...

🔹 Publication Date: Published on Dec 30, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.24176
• PDF: https://arxiv.org/pdf/2512.24176
• Project Page: https://zhouxingyu13.github.io/Internal-Guidance/
• Github: https://github.com/CVL-UESTC/Internal-Guidance

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

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

#DiffusionModels #AI #DeepLearning #GenerativeAI #ComputerVision

✨FlowBlending: Stage-Aware Multi-Model Sampling for Fast and High-Fidelity Video Generation

📝 Summary:
FlowBlending optimizes video generation by adapting model capacity to each stage. It uses large models for critical early and late timesteps, and small models for intermediate ones. This achieves faster inference and fewer FLOPs with no loss in large model fidelity.

🔹 Publication Date: Published on Dec 31, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.24724
• PDF: https://arxiv.org/pdf/2512.24724

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

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

#VideoGeneration #GenerativeAI #DeepLearning #AIResearch #ModelOptimization