✨Assessing Domain-Level Susceptibility to Emergent Misalignment from Narrow Finetuning

📝 Summary:
Large language models fine-tuned on insecure datasets exhibit increased misalignment rates across diverse domains, with varying vulnerability levels and potential for generalization of misalignment be...

🔹 Publication Date: Published on Jan 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.00298
• PDF: https://arxiv.org/pdf/2602.00298

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Reinforced Attention Learning

📝 Summary:
Reinforced Attention Learning optimizes internal attention distributions in multimodal language models, improving information allocation and cross-modal alignment through policy-gradient methods. AI-g...

🔹 Publication Date: Published on Feb 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.04884
• PDF: https://arxiv.org/pdf/2602.04884

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Dr. Kernel: Reinforcement Learning Done Right for Triton Kernel Generations

📝 Summary:
Reinforcement learning approach for kernel generation addresses reward hacking and optimization issues through specialized environment and unbiased policy gradient methods, achieving competitive perfo...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05885
• PDF: https://arxiv.org/pdf/2602.05885

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Do Vision-Language Models Respect Contextual Integrity in Location Disclosure?

📝 Summary:
Vision-language models can precisely geolocate images but often fail to align with human privacy expectations, over-disclosing location details in sensitive contexts and being vulnerable to prompt-bas...

🔹 Publication Date: Published on Feb 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05023
• PDF: https://arxiv.org/pdf/2602.05023
• Project Page: https://huggingface.co/datasets/RayY/VLM-GeoPrivacyBench
• Github: https://github.com/99starman/VLM-GeoPrivacyBench

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨V-Retrver: Evidence-Driven Agentic Reasoning for Universal Multimodal Retrieval

📝 Summary:
V-Retrver introduces an evidence-driven retrieval framework that enables multimodal large language models to actively verify visual evidence through an agentic reasoning process, improving retrieval a...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.06034
• PDF: https://arxiv.org/pdf/2602.06034
• Github: https://github.com/chendy25/V-Retrver

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Spider-Sense: Intrinsic Risk Sensing for Efficient Agent Defense with Hierarchical Adaptive Screening

📝 Summary:
Spider-Sense is an event-driven framework for agent security using Intrinsic Risk Sensing. It provides intrinsic, selective defense through a hierarchical mechanism, activating only upon risk perception. It achieves low attack success and false positive rates with minimal latency.

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05386
• PDF: https://arxiv.org/pdf/2602.05386
• Github: https://github.com/aifinlab/Spider-Sense

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

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

#Cybersecurity #AgentSecurity #AISecurity #RiskSensing #AutonomousAgents

✨Multi-Task GRPO: Reliable LLM Reasoning Across Tasks

📝 Summary:
Multi-Task GRPO MT-GRPO improves LLM reasoning by addressing imbalanced performance across diverse tasks. It dynamically adapts task weights and uses a ratio-preserving sampler to optimize worst-task accuracy. MT-GRPO significantly outperforms baselines in worst-task performance and efficiency.

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05547
• PDF: https://arxiv.org/pdf/2602.05547

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

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

#LLM #ReinforcementLearning #MachineLearning #AI #NLP

✨Light Forcing: Accelerating Autoregressive Video Diffusion via Sparse Attention

📝 Summary:
Light Forcing introduces a sparse attention mechanism for autoregressive video generation. It tackles efficiency bottlenecks using Chunk-Aware Growth and Hierarchical Sparse Attention, improving speed and quality. This method outperforms existing sparse attention, achieving significant speedups.

🔹 Publication Date: Published on Feb 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.04789
• PDF: https://arxiv.org/pdf/2602.04789

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

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

#VideoGeneration #SparseAttention #DeepLearning #AIResearch #ComputerVision

✨Adaptive 1D Video Diffusion Autoencoder

📝 Summary:
One-DVA is a transformer video autoencoder with adaptive encoding and diffusion decoding. It enables variable-length latents and improved compression and detail recovery, addressing fixed-rate compression and deterministic reconstruction.

🔹 Publication Date: Published on Feb 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.04220
• PDF: https://arxiv.org/pdf/2602.04220

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

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

#VideoAI #DiffusionModels #Autoencoders #DeepLearning #ComputerVision

✨PhysicsAgentABM: Physics-Guided Generative Agent-Based Modeling

📝 Summary:
PhysicsAgentABM introduces a neuro-symbolic framework that combines mechanistic agents with neural models to improve scalable and calibrated simulation across multiple domains. AI-generated summary La...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.06030
• PDF: https://arxiv.org/pdf/2602.06030

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Accurate Failure Prediction in Agents Does Not Imply Effective Failure Prevention

📝 Summary:
High offline accuracy in LLM critics does not guarantee effective deployment and can even degrade performance due to a disruption-recovery tradeoff. A small pilot test can predict whether intervention will help or harm, primarily preventing severe regressions before deployment.

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.03338
• PDF: https://arxiv.org/pdf/2602.03338

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

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

#LLM #AISafety #MachineLearning #AIStrategy #Reliability

✨Thinking in Frames: How Visual Context and Test-Time Scaling Empower Video Reasoning

📝 Summary:
Video generation models empower visual reasoning by using generated frames as intermediate steps. They demonstrate robust zero-shot generalization, effectively utilize visual context, and improve planning with increased generated video length.

🔹 Publication Date: Published on Jan 28

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21037
• PDF: https://arxiv.org/pdf/2601.21037
• Github: https://thinking-in-frames.github.io/

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

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

#VideoReasoning #VideoGeneration #ComputerVision #AI #DeepLearning

✨CAR-bench: Evaluating the Consistency and Limit-Awareness of LLM Agents under Real-World Uncertainty

📝 Summary:
CAR-bench evaluates LLM agent reliability under real-world uncertainty, focusing on consistency and capability awareness in in-car assistants. It introduces Hallucination and Disambiguation tasks. Baseline LLMs struggle with disambiguation and often hallucinate, highlighting a need for more relia...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22027
• PDF: https://arxiv.org/pdf/2601.22027
• Github: https://github.com/CAR-bench/car-bench

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

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

#LLMAgents #AI #AutonomousVehicles #AIReliability #AIUncertainty

✨Fast-SAM3D: 3Dfy Anything in Images but Faster

📝 Summary:
Fast-SAM3D addresses slow 3D reconstruction by dynamically adapting computation to varying complexity. It uses heterogeneity-aware mechanisms to achieve up to 2.67x faster inference with negligible quality loss, setting a new efficiency standard.

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05293
• PDF: https://arxiv.org/pdf/2602.05293
• Github: https://github.com/wlfeng0509/Fast-SAM3D

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

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

#3DReconstruction #ComputerVision #DeepLearning #AI #Efficiency

✨Approximation of Log-Partition Function in Policy Mirror Descent Induces Implicit Regularization for LLM Post-Training

📝 Summary:
Policy mirror descent for LLMs struggles with partition function estimation. PMD-mean approximates this with mean reward, implicitly adding a chi-squared regularizer. This enhances robustness and stability, improving LLM post-training performance.

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05933
• PDF: https://arxiv.org/pdf/2602.05933
• Github: https://github.com/horizon-rl/OpenKimi

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

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

#LLM #PolicyMirrorDescent #ReinforcementLearning #MachineLearning #Regularization

✨A Unified Framework for Rethinking Policy Divergence Measures in GRPO

📝 Summary:
This paper presents a unified framework for policy divergence measures in reinforcement learning. It introduces the KL3 estimator as a key constraint, which improves GRPO training stability and performance by promoting stronger exploration.

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05494
• PDF: https://arxiv.org/pdf/2602.05494

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

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

#ReinforcementLearning #MachineLearning #AI #GRPO #PolicyOptimization

✨Focus-dLLM: Accelerating Long-Context Diffusion LLM Inference via Confidence-Guided Context Focusing

📝 Summary:
Focus-dLLM accelerates long-context dLLM inference with a training-free attention sparsification framework. It predicts unmasked regions using confidence-guided indicators and prunes redundant attention while preserving influential sinks across layers. This achieves over 29 times lossless speedup...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02159
• PDF: https://arxiv.org/pdf/2602.02159
• Github: https://github.com/Longxmas/Focus-dLLM

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Towards Reducible Uncertainty Modeling for Reliable Large Language Model Agents

📝 Summary:
Large language models require uncertainty quantification frameworks that account for interactive agent behavior rather than traditional single-turn question answering scenarios. AI-generated summary U...

🔹 Publication Date: Published on Feb 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05073
• PDF: https://arxiv.org/pdf/2602.05073

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Privileged Information Distillation for Language Models

📝 Summary:
This paper introduces pi-Distill and OPSD, methods to distill privileged information PI to language models acting without it. They jointly train a PI-conditioned teacher and unconditioned student. These algorithms effectively transfer PI capabilities, outperforming standard fine-tuning and RL.

🔹 Publication Date: Published on Feb 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.04942
• PDF: https://arxiv.org/pdf/2602.04942

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents

📝 Summary:
MemSkill introduces a learnable and evolvable memory system for LLM agents. It dynamically selects and refines memory operations via a controller, executor, and designer. This closed-loop process improves memory management and outperforms fixed systems.

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02474
• PDF: https://arxiv.org/pdf/2602.02474
• Project Page: https://viktoraxelsen.github.io/MemSkill/
• Github: https://github.com/ViktorAxelsen/MemSkill

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Infinite-World: Scaling Interactive World Models to 1000-Frame Horizons via Pose-Free Hierarchical Memory

📝 Summary:
Infinite-World is a robust interactive world model that maintains coherent visual memory over 1000+ frames through hierarchical pose-free memory compression, uncertainty-aware action labeling, and rev...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02393
• PDF: https://arxiv.org/pdf/2602.02393
• Project Page: https://rq-wu.github.io/projects/infinite-world/index.html

🔹 Models citing this paper:
• https://huggingface.co/MeiGen-AI/Infinite-World

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research