✨Penguin-VL: Exploring the Efficiency Limits of VLM with LLM-based Vision Encoders

📝 Summary:
Penguin-VL introduces a vision encoder initialized from a text-only LLM, outperforming traditional contrastive pretraining. This method achieves superior visual fidelity and performance in multimodal tasks with a lightweight architecture, enabling efficient deployment on resource-constrained devi...

🔹 Publication Date: Published on Mar 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.06569
• PDF: https://arxiv.org/pdf/2603.06569
• Github: https://github.com/tencent-ailab/Penguin-VL

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

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

#VLM #LLM #MultimodalAI #EfficientAI #AIResearch

✨RoboMME: Benchmarking and Understanding Memory for Robotic Generalist Policies

📝 Summary:
RoboMME introduces a large-scale standardized benchmark for evaluating memory in vision-language-action models for long-horizon robotic manipulation. It comprises 16 tasks assessing temporal, spatial, object, and procedural memory. Experiments show memory effectiveness is highly task-dependent, w...

🔹 Publication Date: Published on Mar 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04639
• PDF: https://arxiv.org/pdf/2603.04639
• Project Page: https://robomme.github.io/
• Github: https://github.com/RoboMME/robomme_benchmark

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

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

#Robotics #AI #Benchmark #RoboticManipulation #Memory

✨Reasoning Models Struggle to Control their Chains of Thought

📝 Summary:
Reasoning models exhibit very low control over their Chain-of-Thought steps compared to their final outputs. This low controllability, though poorly understood, currently suggests CoT monitoring remains a reliable tool for understanding models.

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05706
• PDF: https://arxiv.org/pdf/2603.05706

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

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

#AI #MachineLearning #ChainOfThought #LLMs #AIResearch

✨Physical Simulator In-the-Loop Video Generation

📝 Summary:
PSIVG integrates a physical simulator into video diffusion processes to generate physically consistent videos while maintaining visual quality and diversity. AI-generated summary Recent advances in di...

🔹 Publication Date: Published on Mar 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.06408
• PDF: https://arxiv.org/pdf/2603.06408
• Project Page: https://vcai.mpi-inf.mpg.de/projects/PSIVG/

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Dynamic Chunking Diffusion Transformer

📝 Summary:
Dynamic Chunking Diffusion Transformer adapts token sequence length based on image content and diffusion timestep, improving efficiency and performance over fixed-token approaches. AI-generated summar...

🔹 Publication Date: Published on Mar 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.06351
• PDF: https://arxiv.org/pdf/2603.06351

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨π-StepNFT: Wider Space Needs Finer Steps in Online RL for Flow-based VLAs

📝 Summary:
Flow-based VLA models face challenges in online RL. We propose π-StepNFT, a critic-free framework that uses step-wise guidance for wider exploration. It improves generalization and robustness in complex environments, offering a scalable solution.

🔹 Publication Date: Published on Mar 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.02083
• PDF: https://arxiv.org/pdf/2603.02083
• Project Page: https://wangst0181.github.io/pi-StepNFT/
• Github: https://github.com/wangst0181/pi-StepNFT

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

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

#ReinforcementLearning #OnlineRL #MachineLearning #DeepLearning #πStepNFT

✨BandPO: Bridging Trust Regions and Ratio Clipping via Probability-Aware Bounds for LLM Reinforcement Learning

📝 Summary:
BandPO addresses entropy collapse in LLM RL by replacing fixed PPO clipping. It uses Band, a dynamic probability-aware projection operator that prevents suppression of high-advantage actions. This method improves stability and exploration.

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04918
• PDF: https://arxiv.org/pdf/2603.04918
• Github: https://github.com/OpenMOSS/BandPO

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

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

#LLM #ReinforcementLearning #PPO #MachineLearning #AIResearch

✨Progressive Residual Warmup for Language Model Pretraining

📝 Summary:
Progressive Residual Warmup ProRes stabilizes transformer pretraining by gradually activating residual connections layer by layer. This 'early layer learns first' strategy improves convergence speed, generalization, and downstream performance.

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05369
• PDF: https://arxiv.org/pdf/2603.05369
• Github: https://github.com/dandingsky/ProRes

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

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

#LLM #Transformer #DeepLearning #NLP #Pretraining

✨SLER-IR: Spherical Layer-wise Expert Routing for All-in-One Image Restoration

📝 Summary:
SLER-IR is an all-in-one image restoration framework using spherical layer-wise expert routing. It introduces a spherical degradation embedding with contrastive learning for reliable routing and a granularity fusion module for non-uniform degradations. It consistently outperforms state-of-the-art...

🔹 Publication Date: Published on Mar 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05940
• PDF: https://arxiv.org/pdf/2603.05940

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨HiMAP-Travel: Hierarchical Multi-Agent Planning for Long-Horizon Constrained Travel

📝 Summary:
HiMAP-Travel is a hierarchical multi-agent framework that solves long-horizon constrained travel planning. It decomposes tasks into strategic coordination and parallel execution, achieving superior performance over baselines and reducing latency.

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04750
• PDF: https://arxiv.org/pdf/2603.04750

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Planning in 8 Tokens: A Compact Discrete Tokenizer for Latent World Model

📝 Summary:
CompACT, a discrete tokenizer that reduces observation encoding from hundreds to 8 tokens, enables faster and more efficient world model planning for real-time control applications. AI-generated summa...

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05438
• PDF: https://arxiv.org/pdf/2603.05438
• Github: https://github.com/kdwonn/CompACT

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨WildActor: Unconstrained Identity-Preserving Video Generation

📝 Summary:
WildActor generates consistent human videos with full-body identity preservation across varying viewpoints and motions using a large-scale dataset and novel attention mechanisms. AI-generated summary ...

🔹 Publication Date: Published on Feb 28

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.00586
• PDF: https://arxiv.org/pdf/2603.00586
• Project Page: https://wildactor.github.io/

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Beyond the Grid: Layout-Informed Multi-Vector Retrieval with Parsed Visual Document Representations

📝 Summary:
ColParse introduces a document parsing approach that generates layout-informed sub-image embeddings to create compact, structurally-aware representations for visual document retrieval, achieving over ...

🔹 Publication Date: Published on Mar 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.01666
• PDF: https://arxiv.org/pdf/2603.01666

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Making Reconstruction FID Predictive of Diffusion Generation FID

📝 Summary:
A new metric called interpolated FID is proposed that shows strong correlation with generation FID in diffusion models, addressing the poor correlation issue between reconstruction FID and generation ...

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05630
• PDF: https://arxiv.org/pdf/2603.05630
• Github: https://github.com/tongdaxu/Making-rFID-Predictive-of-Diffusion-gFID

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Demystifying Action Space Design for Robotic Manipulation Policies

📝 Summary:
Large-scale empirical study demonstrates that action space design significantly impacts robotic policy learning, with delta action prediction improving performance and joint-space/task-space represent...

🔹 Publication Date: Published on Feb 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.23408
• PDF: https://arxiv.org/pdf/2602.23408

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Mario: Multimodal Graph Reasoning with Large Language Models

📝 Summary:
Mario is a unified framework that enables large language model-based reasoning on multimodal graphs by addressing cross-modal consistency and heterogeneous modality preferences through graph-condition...

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05181
• PDF: https://arxiv.org/pdf/2603.05181
• Github: https://github.com/sunyuanfu/Mario

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨DeepPresenter: Environment-Grounded Reflection for Agentic Presentation Generation

📝 Summary:
DeepPresenter is an agentic framework for adaptive presentation generation. It plans and refines slide artifacts using environment-grounded reflection on rendered slides. This approach achieves state-of-the-art performance with reduced computational costs.

🔹 Publication Date: Published on Feb 26

🔹 Paper Links:
• arXiv Page: https://huggingface.co/collections/ICIP/deeppresenter
• PDF: https://arxiv.org/pdf/2602.22839
• Github: https://github.com/icip-cas/PPTAgent

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

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

#AI #AgenticAI #PresentationGeneration #DeepLearning #GenerativeAI

✨WorldCache: Accelerating World Models for Free via Heterogeneous Token Caching

📝 Summary:
WorldCache speeds up slow diffusion-based world models by addressing token heterogeneity and non-uniform dynamics. It uses curvature-guided prediction and chaotic-prioritized skipping. This achieves up to 3.7 times faster inference with 98 percent rollout quality.

🔹 Publication Date: Published on Mar 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.06331
• PDF: https://arxiv.org/pdf/2603.06331

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

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

#WorldModels #DiffusionModels #AI #MachineLearning #Optimization

✨Layer by layer, module by module: Choose both for optimal OOD probing of ViT

📝 Summary:
Intermediate layers in ViTs provide better representations. Performance degradation in deeper layers is caused by distribution shift. Optimal probing depends on shift magnitude: FFN activation for strong shift, MHA output for weak shift.

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05280
• PDF: https://arxiv.org/pdf/2603.05280
• Github: https://github.com/ambroiseodt/vit-probing

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

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

#ViT #OOD #DeepLearning #RepresentationLearning #ComputerVision

✨Dynamic Model Routing and Cascading for Efficient LLM Inference: A Survey

📝 Summary:
This survey analyzes dynamic routing systems that adaptively select among multiple independent LLMs based on query characteristics to optimize inference performance and cost. It covers diverse routing paradigms and presents a framework for understanding these systems, highlighting their ability t...

🔹 Publication Date: Published on Feb 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04445
• PDF: https://arxiv.org/pdf/2603.04445

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

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

#LLM #AI #ModelRouting #InferenceOptimization #DeepLearning

✨EffectMaker: Unifying Reasoning and Generation for Customized Visual Effect Creation

📝 Summary:
EffectMaker is a unified framework for reference-based VFX customization. It uses a multimodal language model and diffusion transformer for semantic-visual guidance, generating high-quality effects consistently without per-effect fine-tuning. This is supported by a large synthetic dataset.

🔹 Publication Date: Published on Mar 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.06014
• PDF: https://arxiv.org/pdf/2603.06014
• Project Page: https://effectmaker.github.io/

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

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

#VFX #GenerativeAI #DiffusionModels #MultimodalAI #ComputerVision