✨Reward Forcing: Efficient Streaming Video Generation with Rewarded Distribution Matching Distillation

📝 Summary:
Reward Forcing improves streaming video generation by using EMA-Sink to update context tokens, preventing static initial frames. It also introduces Rewarded Distribution Matching Distillation to prioritize dynamic content, enhancing motion quality and achieving state-of-the-art performance.

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.04678
• PDF: https://arxiv.org/pdf/2512.04678
• Project Page: https://reward-forcing.github.io/
• Github: https://reward-forcing.github.io/

🔹 Models citing this paper:
• https://huggingface.co/JaydenLu666/Reward-Forcing-T2V-1.3B

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#VideoGeneration #GenerativeAI #DeepLearning #ComputerVision #AIResearch

✨SeeNav-Agent: Enhancing Vision-Language Navigation with Visual Prompt and Step-Level Policy Optimization

📝 Summary:
SeeNav-Agent improves Vision-Language Navigation with dual-view visual prompts, reducing perception errors and enhancing spatial understanding. It also uses SRGPO, a step-level reinforcement fine-tuning method, to boost planning and achieve higher success rates for VLN agents.

🔹 Publication Date: Published on Dec 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.02631
• PDF: https://arxiv.org/pdf/2512.02631
• Github: https://github.com/WzcTHU/SeeNav-Agent

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#VisionLanguageNavigation #AI #ReinforcementLearning #ComputerVision #DeepLearning

✨Splannequin: Freezing Monocular Mannequin-Challenge Footage with Dual-Detection Splatting

📝 Summary:
Splannequin improves frozen 3D scenes from monocular videos by fixing artifacts in dynamic Gaussian splatting. It uses temporal anchoring for hidden or defective Gaussians to resolve ghosting and blur from sparse supervision. This boosts visual quality for high-fidelity, user-selectable frozen-ti...

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.05113
• PDF: https://arxiv.org/pdf/2512.05113
• Project Page: https://chien90190.github.io/splannequin/

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#ComputerVision #3DReconstruction #GaussianSplatting #NeuralRendering #DeepLearning

✨Nex-N1: Agentic Models Trained via a Unified Ecosystem for Large-Scale Environment Construction

📝 Summary:
Training autonomous LLM agents requires scalable, high-quality interactive environments. The Nex ecosystem provides NexAU for complexity, NexA4A for diversity, and NexGAP for fidelity in environment construction. Nex-N1, trained using this infrastructure, outperforms SOTA models on agentic tasks.

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.04987
• PDF: https://arxiv.org/pdf/2512.04987
• Github: https://github.com/nex-agi/Nex-N1

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#LLMAgents #LargeLanguageModels #AI #AISimulation #AIResearch

✨Semantics Lead the Way: Harmonizing Semantic and Texture Modeling with Asynchronous Latent Diffusion

📝 Summary:
Semantic-First Diffusion SFD asynchronously denoises semantic and texture latents for image generation. This method prioritizes semantic formation, providing clearer guidance for texture refinement. SFD significantly improves convergence speed by up to 100x and enhances image quality.

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.04926
• PDF: https://arxiv.org/pdf/2512.04926

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#DiffusionModels #ImageGeneration #SemanticAI #GenerativeAI #DeepLearning

✨SignRoundV2: Closing the Performance Gap in Extremely Low-Bit Post-Training Quantization for LLMs

📝 Summary:
SignRoundV2 is a post-training quantization framework for LLMs. It uses a sensitivity metric for bit allocation and pre-tuning for scales to achieve competitive accuracy even at 2-bit quantization, closing the gap with full-precision models.

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.04746
• PDF: https://arxiv.org/pdf/2512.04746

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#LLMs #Quantization #DeepLearning #AI #MachineLearning

✨TV2TV: A Unified Framework for Interleaved Language and Video Generation

📝 Summary:
TV2TV is a unified framework for interleaved language and video generation, using a Mixture-of-Transformers. It learns to 'think in words' before 'acting in pixels,' enhancing visual quality, controllability, and prompt alignment. The model shows strong performance on video game and natural video...

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.05103
• PDF: https://arxiv.org/pdf/2512.05103

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#VideoGeneration #GenerativeAI #MultimodalAI #Transformers #AI

✨DAComp: Benchmarking Data Agents across the Full Data Intelligence Lifecycle

📝 Summary:
DAComp is a benchmark with 210 tasks for data engineering and analysis workflows. It reveals significant deficiencies in state-of-the-art agents, with success rates under 20% for engineering and below 40% for analysis, highlighting critical gaps.

🔹 Publication Date: Published on Dec 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.04324
• PDF: https://arxiv.org/pdf/2512.04324
• Project Page: https://da-comp.github.io/

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#DataAgents #Benchmarking #DataEngineering #DataAnalysis #AIResearch

✨On GRPO Collapse in Search-R1: The Lazy Likelihood-Displacement Death Spiral

📝 Summary:
GRPO in tool-integrated RL collapses due to Lazy Likelihood Displacement LLD, a systematic drop in response likelihoods. LLDS regularization addresses this by preserving likelihoods, stabilizing training, preventing gradient explosion, and substantially improving performance.

🔹 Publication Date: Published on Dec 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.04220
• PDF: https://arxiv.org/pdf/2512.04220

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#ReinforcementLearning #MachineLearning #AI #DeepLearning #AIResearch

✨Stable Video Infinity: Infinite-Length Video Generation with Error Recycling

📝 Summary:
Stable Video Infinity SVI generates infinite-length videos with high consistency and controllable stories. It introduces Error-Recycling Fine-Tuning, teaching the Diffusion Transformer to correct its self-generated errors and address the training-test discrepancy.

🔹 Publication Date: Published on Oct 10

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.09212
• PDF: https://arxiv.org/pdf/2510.09212
• Project Page: https://stable-video-infinity.github.io/homepage/
• Github: https://github.com/vita-epfl/Stable-Video-Infinity

🔹 Models citing this paper:
• https://huggingface.co/vita-video-gen/svi-model

✨ Datasets citing this paper:
• https://huggingface.co/datasets/vita-video-gen/svi-benchmark

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#VideoGeneration #AI #DiffusionModels #DeepLearning #ComputerVision

✨PaperDebugger: A Plugin-Based Multi-Agent System for In-Editor Academic Writing, Review, and Editing

📝 Summary:
PaperDebugger is an in-editor, multi-agent academic writing assistant that integrates large language models directly into LaTeX environments. It allows deep interaction with document state and revision history for enhanced writing, review, and editing workflows.

🔹 Publication Date: Published on Dec 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.02589
• PDF: https://arxiv.org/pdf/2512.02589
• Project Page: https://www.paperdebugger.com/
• Github: https://github.com/PaperDebugger/PaperDebugger

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#AcademicWriting #LLM #MultiAgentSystems #ResearchTools #AI

✨DynamicVerse: A Physically-Aware Multimodal Framework for 4D World Modeling

📝 Summary:
DynamicVerse introduces a 4D world modeling framework for dynamic real-world videos, overcoming existing dataset limitations. It integrates large vision, geometric, and multimodal models to create a vast dataset with metric-scale annotations. This approach achieves superior performance in depth, ...

🔹 Publication Date: Published on Dec 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.03000
• PDF: https://arxiv.org/pdf/2512.03000
• Project Page: https://dynamic-verse.github.io/
• Github: https://dynamic-verse.github.io/

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#4DModeling #MultimodalAI #ComputerVision #DeepLearning #AIResearch

✨DraCo: Draft as CoT for Text-to-Image Preview and Rare Concept Generation

📝 Summary:
DraCo is a novel text-to-image generation method that uses interleaved reasoning with both textual and visual content. It generates low-resolution drafts, verifies semantic alignment, and refines images to address coarse textual planning and rare attribute generation. DraCo significantly outperfo...

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.05112
• PDF: https://arxiv.org/pdf/2512.05112
• Github: https://github.com/CaraJ7/DraCo

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#TextToImage #GenerativeAI #DeepLearning #ComputerVision #AI

✨BulletTime: Decoupled Control of Time and Camera Pose for Video Generation

📝 Summary:
This paper presents a video diffusion framework that decouples scene dynamics from camera pose. This enables precise 4D control over time and viewpoint for high-quality video generation, outperforming prior models in controllability.

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.05076
• PDF: https://arxiv.org/pdf/2512.05076

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#VideoGeneration #DiffusionModels #GenerativeAI #ComputerVision #AICameraControl

✨Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length

📝 Summary:
Live Avatar uses a 14-billion-parameter diffusion model to achieve real-time, high-fidelity, infinite-length audio-driven avatar generation. It employs Timestep-forcing Pipeline Parallelism and Rolling Sink Frame Mechanism for efficiency and consistency, reaching 20 FPS on 5 H800 GPUs.

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.04677
• PDF: https://arxiv.org/pdf/2512.04677

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#LiveAvatar #GenerativeAI #RealtimeAI #DiffusionModels #AvatarGeneration

✨NeuralRemaster: Phase-Preserving Diffusion for Structure-Aligned Generation

📝 Summary:
Standard diffusion corrupts image phase, destroying spatial structure. This paper introduces Phase-Preserving Diffusion phi-PD to preserve phase, enabling structure-aligned generation for tasks like re-rendering. It adds no cost and improves sim-to-real enhancement significantly.

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.05106
• PDF: https://arxiv.org/pdf/2512.05106

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#DiffusionModels #GenerativeAI #ComputerVision #DeepLearning #AIResearch

✨REFLEX: Self-Refining Explainable Fact-Checking via Disentangling Truth into Style and Substance

📝 Summary:
REFLEX is a new fact-checking method that uses internal model knowledge to improve verdict accuracy and explanation quality. It disentangles truth into style and substance via adaptive activation signals, achieving state-of-the-art performance with minimal training data. This approach also shows ...

🔹 Publication Date: Published on Nov 25

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2511.20233
• PDF: https://arxiv.org/pdf/2511.20233

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#FactChecking #ExplainableAI #MachineLearning #AI #NLP

✨EgoLCD: Egocentric Video Generation with Long Context Diffusion

📝 Summary:
EgoLCD addresses content drift in long egocentric video generation by integrating long-term sparse and attention-based short-term memory with narrative prompting. It achieves state-of-the-art perceptual quality and temporal consistency, mitigating generative forgetting.

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.04515
• PDF: https://arxiv.org/pdf/2512.04515
• Project Page: https://aigeeksgroup.github.io/EgoLCD/
• Github: https://github.com/AIGeeksGroup/EgoLCD

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#AI #VideoGeneration #DiffusionModels #ComputerVision #EgocentricVision

✨ARM-Thinker: Reinforcing Multimodal Generative Reward Models with Agentic Tool Use and Visual Reasoning

📝 Summary:
ARM-Thinker is an agentic reward model that uses external tools like image cropping and document retrieval to verify judgments in multimodal reasoning tasks. This significantly improves accuracy, interpretability, and visual grounding compared to existing reward models, achieving substantial perf...

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.05111
• PDF: https://arxiv.org/pdf/2512.05111
• Project Page: https://github.com/InternLM/ARM-Thinker
• Github: https://github.com/open-compass/VLMEvalKit/pull/1334

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#MultimodalAI #AgenticAI #RewardModels #VisualReasoning #AIResearch

✨QKAN-LSTM: Quantum-inspired Kolmogorov-Arnold Long Short-term Memory

📝 Summary:
QKAN-LSTM is a quantum-inspired LSTM that integrates Data Re-Uploading Activation modules. This model achieves superior predictive accuracy and generalization with 79% fewer parameters than classical LSTMs. It offers a scalable, interpretable approach for sequential modeling on classical hardware.

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.05049
• PDF: https://arxiv.org/pdf/2512.05049

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#QKANLSTM #QuantumInspiredAI #DeepLearning #MachineLearning #DataScience

✨Mitigating Object and Action Hallucinations in Multimodal LLMs via Self-Augmented Contrastive Alignment

📝 Summary:
The SANTA framework addresses object and action hallucinations in multimodal LLM video captions. It uses self-augmented contrastive alignment to identify potential hallucinations and then aligns regional objects and actions with visual phrases, improving factual accuracy. Experiments show SANTA o...

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.04356
• PDF: https://arxiv.org/pdf/2512.04356
• Project Page: https://kpc0810.github.io/santa/
• Github: https://kpc0810.github.io/santa/

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

For more data science resources:
✓ https://news.1rj.ru/str/DataScienceT

#MultimodalLLMs #AI #Hallucinations #VideoUnderstanding #ContrastiveLearning