✨VESPO: Variational Sequence-Level Soft Policy Optimization for Stable Off-Policy LLM Training

📝 Summary:
VESPO addresses LLM RL training instability by using a variational formulation with variance reduction. It provides a sequence-level correction without length normalization, ensuring stable training and consistent gains even with high policy staleness.

🔹 Publication Date: Published on Feb 11

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.10693
• PDF: https://arxiv.org/pdf/2602.10693
• Github: https://github.com/FloyedShen/VESPO

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

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

#LLM #ReinforcementLearning #DeepLearning #AI #NLP

✨AudioX: Diffusion Transformer for Anything-to-Audio Generation

📝 Summary:
AudioX is a unified Diffusion Transformer for high-quality audio and music generation with natural language control. It processes diverse modalities using a novel multi-modal masked training strategy. This model outperforms specialized systems while offering remarkable versatility.

🔹 Publication Date: Published on Mar 13, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2503.10522
• PDF: https://arxiv.org/pdf/2503.10522
• Project Page: https://zeyuet.github.io/AudioX/
• Github: https://github.com/ZeyueT/AudioX

🔹 Models citing this paper:
• https://huggingface.co/HKUSTAudio/AudioX
• https://huggingface.co/HKUSTAudio/AudioX-MAF-MMDiT
• https://huggingface.co/Zeyue7/AudioX

✨ Datasets citing this paper:
• https://huggingface.co/datasets/HKUSTAudio/AudioX-IFcaps

✨ Spaces citing this paper:
• https://huggingface.co/spaces/Zeyue7/AudioX
• https://huggingface.co/spaces/Napawit/AudioX
• https://huggingface.co/spaces/ar93092/atai

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

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

#AudioGeneration #DiffusionModels #Transformers #AI #MultimodalAI

✨Selective Training for Large Vision Language Models via Visual Information Gain

📝 Summary:
This paper proposes Visual Information Gain VIG to quantify visual inputs contribution to prediction uncertainty in Large Vision Language Models. VIG enables selective training, improving visual grounding and reducing language bias with less supervision.

🔹 Publication Date: Published on Feb 19

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.17186
• PDF: https://arxiv.org/pdf/2602.17186

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

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

#LVLMs #SelectiveTraining #VisualInformationGain #ComputerVision #AIResearch

✨DeepVision-103K: A Visually Diverse, Broad-Coverage, and Verifiable Mathematical Dataset for Multimodal Reasoning

📝 Summary:
To address limitations in existing datasets, DeepVision-103K offers a comprehensive and visually diverse mathematical dataset for multimodal reasoning. It enhances model performance, visual perception, and reasoning in large multimodal models.

🔹 Publication Date: Published on Feb 18

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.16742
• PDF: https://arxiv.org/pdf/2602.16742
• Github: https://github.com/SKYLENAGE-AI/DeepVision-103K

✨ Datasets citing this paper:
• https://huggingface.co/datasets/skylenage/DeepVision-103K

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

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

#MultimodalAI #ComputerVision #Datasets #AIResearch #DeepLearning

✨Mobile-Agent-v3: Foundamental Agents for GUI Automation

📝 Summary:
This paper introduces GUI-Owl and Mobile-Agent-v3, open-source GUI agent models and frameworks. Mobile-Agent-v3 achieves new state-of-the-art performance on GUI automation benchmarks like AndroidWorld and OSWorld by building on GUI-Owl's innovations in environment infrastructure, agent capabiliti...

🔹 Publication Date: Published on Aug 21, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2508.15144
• PDF: https://arxiv.org/pdf/2508.15144
• Project Page: https://github.com/X-PLUG/MobileAgent
• Github: https://github.com/X-PLUG/MobileAgent

🔹 Models citing this paper:
• https://huggingface.co/mPLUG/GUI-Owl-7B
• https://huggingface.co/mPLUG/GUI-Owl-32B
• https://huggingface.co/mPLUG/GUI-Owl-7B-Desktop-RL

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

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

#GUIAgent #Automation #AI #OpenSource #MachineLearning

✨VidEoMT: Your ViT is Secretly Also a Video Segmentation Model

📝 Summary:
VidEoMT is a video segmentation model that eliminates complex tracking modules by using a Vision Transformer encoder with query propagation and fusion. This enables efficient temporal modeling, achieving competitive accuracy and 5-10x faster processing speeds.

🔹 Publication Date: Published on Feb 19

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.17807
• PDF: https://arxiv.org/pdf/2602.17807
• Project Page: https://www.tue-mps.org/videomt/
• Github: https://github.com/tue-mps/videomt

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

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

#VideoSegmentation #VisionTransformers #ComputerVision #DeepLearning #AIResearch

✨Sink-Aware Pruning for Diffusion Language Models

📝 Summary:
Diffusion Language Models have high inference costs. This paper finds that their attention sinks are often unstable, unlike in autoregressive models. Sink-Aware Pruning identifies and removes these unstable sinks, improving efficiency and quality without retraining.

🔹 Publication Date: Published on Feb 19

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.17664
• PDF: https://arxiv.org/pdf/2602.17664
• Github: https://github.com/VILA-Lab/Sink-Aware-Pruning

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

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

#DiffusionModels #LanguageModels #ModelPruning #NLP #AIResearch

✨PersonaLive! Expressive Portrait Image Animation for Live Streaming

📝 Summary:
PersonaLive enables real-time, expressive portrait animation for live streaming. It uses hybrid implicit signals, appearance distillation, and autoregressive streaming generation to achieve low-latency, stable results with up to 22x speedup.

🔹 Publication Date: Published on Dec 12, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.11253
• PDF: https://arxiv.org/pdf/2512.11253
• Github: https://github.com/GVCLab/PersonaLive

🔹 Models citing this paper:
• https://huggingface.co/huaichang/PersonaLive

✨ Spaces citing this paper:
• https://huggingface.co/spaces/seawolf2357/personalive

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

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

#PortraitAnimation #LiveStreaming #RealtimeAI #ComputerVision #GenerativeAI

✨Decoding as Optimisation on the Probability Simplex: From Top-K to Top-P (Nucleus) to Best-of-K Samplers

📝 Summary:
This paper redefines decoding as an optimization problem on the probability simplex balancing model scores with structural preferences. This unifies existing methods and enables new decoders like Best-of-K, improving accuracy in tasks such as mathematical reasoning.

🔹 Publication Date: Published on Feb 20

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.18292
• PDF: https://arxiv.org/pdf/2602.18292

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

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

#DecodingStrategies #Optimization #LLMs #MathematicalReasoning #MachineLearning

✨4RC: 4D Reconstruction via Conditional Querying Anytime and Anywhere

📝 Summary:
4RC introduces a unified feed-forward framework for 4D reconstruction from monocular video. It learns holistic scene geometry and motion dynamics using a novel transformer-based 'encode-once, query-anywhere and anytime' approach. This method significantly outperforms prior 4D reconstruction techn...

🔹 Publication Date: Published on Feb 10

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.10094
• PDF: https://arxiv.org/pdf/2602.10094
• Project Page: https://yihangluo.com/projects/4RC/

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

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

#4DReconstruction #ComputerVision #DeepLearning #NeuralNetworks #MonocularVideo

✨Spanning the Visual Analogy Space with a Weight Basis of LoRAs

📝 Summary:
LoRWeB improves visual analogy learning by dynamically composing a basis of LoRA modules. It uses an encoder to select and weigh multiple LoRAs at inference time, rather than a single fixed module. This achieves state-of-the-art performance and significantly better generalization for image manipu...

🔹 Publication Date: Published on Feb 17

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.15727
• PDF: https://arxiv.org/pdf/2602.15727
• Project Page: https://research.nvidia.com/labs/par/lorweb/
• Github: https://github.com/NVlabs/LoRWeB

✨ Datasets citing this paper:
• https://huggingface.co/datasets/hilamanor/LoRWeB_evalset

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

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

#LoRA #VisualAnalogies #DeepLearning #AI #ComputerVision

✨Adam Improves Muon: Adaptive Moment Estimation with Orthogonalized Momentum

📝 Summary:
NAMO and NAMO-D are new optimizers combining orthogonalized momentum with Adam-type noise adaptation. They show improved convergence and better performance on LLM pretraining than AdamW and Muon, with NAMO-D adding neuron-wise adaptation for further gains.

🔹 Publication Date: Published on Feb 19

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.17080
• PDF: https://arxiv.org/pdf/2602.17080

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

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

#MachineLearning #DeepLearning #LLM #Optimizers #Adam

✨Avey-B

📝 Summary:
This paper reformulates the Avey architecture for encoder-only tasks, introducing innovations like decoupled parameterizations and neural compression. The new model consistently outperforms Transformer-based encoders on token classification and information retrieval, also scaling more efficiently...

🔹 Publication Date: Published on Feb 17

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.15814
• PDF: https://arxiv.org/pdf/2602.15814
• Github: https://github.com/rimads/avey-b

🔹 Models citing this paper:
• https://huggingface.co/avey-ai/avey-b1-base-exp
• https://huggingface.co/avey-ai/avey-b1-large-exp

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨ReIn: Conversational Error Recovery with Reasoning Inception

📝 Summary:
Conversational agents with tool integration face challenges from user-induced errors, but a test-time intervention method called Reasoning Inception (ReIn) enables error recovery by injecting external...

🔹 Publication Date: Published on Feb 19

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.17022
• PDF: https://arxiv.org/pdf/2602.17022

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Whom to Query for What: Adaptive Group Elicitation via Multi-Turn LLM Interactions

📝 Summary:
An adaptive group elicitation framework combines LLM information gain with graph neural networks for population predictions. It selects questions and respondents, imputing missing data under budget limits to improve prediction accuracy with fewer queries.

🔹 Publication Date: Published on Feb 15

🔹 Paper Links:
• arXiv Page: https://arxiv.org/pdf/2602.14279
• PDF: https://arxiv.org/pdf/2602.14279
• Github: https://github.com/ZDCSlab/Group-Adaptive-Elicitation

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Rubrics as an Attack Surface: Stealthy Preference Drift in LLM Judges

📝 Summary:
LLM-based judges using natural-language rubrics for evaluation can exhibit systematic preference drift from minor rubric modifications, which can be exploited to manipulate alignment pipelines and deg...

🔹 Publication Date: Published on Feb 14

🔹 Paper Links:
• arXiv Page: https://arxiv.org/pdf/2602.13576
• PDF: https://arxiv.org/pdf/2602.13576
• Github: https://github.com/ZDCSlab/Rubrics-as-an-Attack-Surface

🔹 Models citing this paper:
• https://huggingface.co/ZDCSlab/ripd-ultra-real-gemma2-2b-it-seed-bt
• https://huggingface.co/ZDCSlab/ripd-ultra-real-gemma2-2b-it-biased-bt
• https://huggingface.co/ZDCSlab/ripd-ultra-real-llama3-8b-instruct-seed-bt

✨ Datasets citing this paper:
• https://huggingface.co/datasets/ZDCSlab/ripd-dataset

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics

📝 Summary:
TOPReward is a novel temporal value function that estimates robotic task progress using pretrained video VLM internal token logits. It achieves superior zero-shot performance across over 130 real-world tasks and multiple robots, greatly outperforming baselines.

🔹 Publication Date: Published on Feb 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.19313
• PDF: https://arxiv.org/pdf/2602.19313
• Project Page: https://topreward.github.io/webpage/
• Github: https://github.com/TOPReward/TOPReward

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Mobile-O: Unified Multimodal Understanding and Generation on Mobile Device

📝 Summary:
A compact vision-language-diffusion model called Mobile-O enables efficient unified multimodal understanding and generation on mobile devices through specialized architecture design and optimized trai...

🔹 Publication Date: Published on Feb 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.20161
• PDF: https://arxiv.org/pdf/2602.20161
• Project Page: https://amshaker.github.io/Mobile-O/
• Github: https://github.com/Amshaker/Mobile-O

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨DSDR: Dual-Scale Diversity Regularization for Exploration in LLM Reasoning

📝 Summary:
DSDR is a reinforcement learning framework that enhances large language model reasoning by promoting diversity at both global and local levels through dual-scale regularization techniques. AI-generate...

🔹 Publication Date: Published on Feb 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.19895
• PDF: https://arxiv.org/pdf/2602.19895

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Agents of Chaos

📝 Summary:
Autonomous language-model-powered agents in a live laboratory environment exhibited numerous security and governance vulnerabilities including unauthorized actions, information disclosure, and system ...

🔹 Publication Date: Published on Feb 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.20021
• PDF: https://arxiv.org/pdf/2602.20021

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research