✨Quartet II: Accurate LLM Pre-Training in NVFP4 by Improved Unbiased Gradient Estimation

📝 Summary:
Quartet II improves LLM pre-training in NVFP4 by introducing MS-EDEN for enhanced unbiased gradient estimation, significantly reducing quantization error. This achieves better accuracy and up to 4.2x faster execution on NVIDIA Blackwell GPUs compared to BF16.

🔹 Publication Date: Published on Jan 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22813
• PDF: https://arxiv.org/pdf/2601.22813
• Github: https://github.com/IST-DASLab/Quartet-II

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

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

#LLM #DeepLearning #Quantization #GPUAcceleration #AIResearch

✨ExpAlign: Expectation-Guided Vision-Language Alignment for Open-Vocabulary Grounding

📝 Summary:
ExpAlign proposes an expectation-guided vision-language alignment framework using multiple instance learning and attention pooling. It implicitly selects tokens and instances without extra annotations, significantly boosting open-vocabulary detection and zero-shot instance segmentation.

🔹 Publication Date: Published on Jan 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22666
• PDF: https://arxiv.org/pdf/2601.22666

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

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

#ComputerVision #DeepLearning #AI #VisionLanguage #OpenVocabulary

✨KAPSO: A Knowledge-grounded framework for Autonomous Program Synthesis and Optimization

📝 Summary:
KAPSO is a modular framework for autonomous program synthesis. It uses iterative optimization loops, a git-native experimentation engine, a comprehensive knowledge system, and cognitive memory to improve code over extended tasks, overcoming common coding agent failures.

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21526
• PDF: https://arxiv.org/pdf/2601.21526
• Github: https://github.com/Leeroo-AI/kapso

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

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

#ProgramSynthesis #AI #CodeOptimization #KnowledgeAI #AIforCoding

✨Do Reasoning Models Enhance Embedding Models?

📝 Summary:
Embedding models from RLVR-tuned reasoning backbones show no performance advantage. HRSA explains this: RLVR reorganizes local geometry but preserves global geometry and linear readout, allowing manifold realignment during contrastive training.

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21192
• PDF: https://arxiv.org/pdf/2601.21192
• Github: https://github.com/HKUST-KnowComp/Reasoning-Embedding

🔹 Models citing this paper:
• https://huggingface.co/lucaswychan/Qwen2.5-1.5B-Reasoning-Embedding
• https://huggingface.co/lucaswychan/Qwen-2.5-1.5B-SimpleRL-Zoo-Reasoning-Embedding
• https://huggingface.co/lucaswychan/Qwen2.5-0.5B-Reasoning-Embedding

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Causal World Modeling for Robot Control

📝 Summary:
Video world modeling enables robot learning through a unified framework that predicts frames and executes policies simultaneously using a shared latent space and closed-loop feedback mechanisms. AI-ge...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21998
• PDF: https://arxiv.org/pdf/2601.21998

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Visual Personalization Turing Test

📝 Summary:
A new evaluation framework called VPTT assesses contextual visual personalization through perceptual indistinguishability from human-created content, utilizing a benchmark, retrieval-augmented generat...

🔹 Publication Date: Published on Jan 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22680
• PDF: https://arxiv.org/pdf/2601.22680

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨SONIC-O1: A Real-World Benchmark for Evaluating Multimodal Large Language Models on Audio-Video Understanding

📝 Summary:
A comprehensive benchmark for evaluating multimodal large language models on sequential audio-video data across real-world conversational domains with human-verified annotations and demographic metada...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21666
• PDF: https://arxiv.org/pdf/2601.21666

✨ Datasets citing this paper:
• https://huggingface.co/datasets/vector-institute/sonic-o1

✨ Spaces citing this paper:
• https://huggingface.co/spaces/vector-institute/sonic-o1-leaderboard

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Value-Based Pre-Training with Downstream Feedback

📝 Summary:
V-Pretraining reshapes foundation model pretraining objectives by using downstream task gradients. This method improves model capabilities and efficiency for tasks like language reasoning and vision segmentation, using minimal downstream feedback without direct label updates.

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22108
• PDF: https://arxiv.org/pdf/2601.22108

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Drive-JEPA: Video JEPA Meets Multimodal Trajectory Distillation for End-to-End Driving

📝 Summary:
Drive-JEPA combines V-JEPA video pretraining with multimodal trajectory distillation to achieve state-of-the-art performance in end-to-end autonomous driving. AI-generated summary End-to-end autonomou...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22032
• PDF: https://arxiv.org/pdf/2601.22032
• Project Page: https://github.com/linhanwang/Drive-JEPA
• Github: https://github.com/linhanwang/Drive-JEPA

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Memorization Dynamics in Knowledge Distillation for Language Models

📝 Summary:
Knowledge distillation reduces training data memorization compared to standard fine-tuning while maintaining performance, with distinct memorization patterns and predictability based on input characte...

🔹 Publication Date: Published on Jan 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.15394
• PDF: https://arxiv.org/pdf/2601.15394

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨RAPTOR: Ridge-Adaptive Logistic Probes

📝 Summary:
RAPTOR is a ridge-adaptive logistic probe that accurately and stably estimates concept vectors for activation steering in frozen LLMs. It significantly reduces training costs while matching or exceeding baseline accuracy and stability. Theoretical analysis underpins its efficacy.

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.00158
• PDF: https://arxiv.org/pdf/2602.00158

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨FS-Researcher: Test-Time Scaling for Long-Horizon Research Tasks with File-System-Based Agents

📝 Summary:
FS-Researcher is a dual-agent framework that scales LLM research tasks beyond context window limits. It uses a file system as persistent external memory, enabling a Context Builder and Report Writer to achieve state-of-the-art report quality and effective test-time scaling.

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.01566
• PDF: https://arxiv.org/pdf/2602.01566

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨How Well Do Models Follow Visual Instructions? VIBE: A Systematic Benchmark for Visual Instruction-Driven Image Editing

📝 Summary:
Visual Instruction Benchmark for Image Editing introduces a three-level interaction hierarchy for evaluating visual instruction following capabilities in generative models. AI-generated summary Recent...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.01851
• PDF: https://arxiv.org/pdf/2602.01851
• Github: https://vibe-benchmark.github.io/

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Ebisu: Benchmarking Large Language Models in Japanese Finance

📝 Summary:
A Japanese financial language understanding benchmark named Ebisu is introduced, featuring two expert-annotated tasks that evaluate implicit commitment recognition and hierarchical financial terminolo...

🔹 Publication Date: Published on Feb 1

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.01479
• PDF: https://arxiv.org/pdf/2602.01479

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨PISCES: Annotation-free Text-to-Video Post-Training via Optimal Transport-Aligned Rewards

📝 Summary:
PISCES is an annotation-free text-to-video generation method that uses dual optimal transport-aligned rewards to improve visual quality and semantic alignment without human preference annotations. AI-...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.01624
• PDF: https://arxiv.org/pdf/2602.01624

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨PromptRL: Prompt Matters in RL for Flow-Based Image Generation

📝 Summary:
Flow matching models for text-to-image generation are enhanced through a reinforcement learning framework that addresses sample inefficiency and prompt overfitting by incorporating language models for...

🔹 Publication Date: Published on Feb 1

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.01382
• PDF: https://arxiv.org/pdf/2602.01382
• Github: https://github.com/G-U-N/UniRL

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Adaptive Ability Decomposing for Unlocking Large Reasoning Model Effective Reinforcement Learning

📝 Summary:
Adaptive Ability Decomposing (A²D) enhances reinforcement learning with verifiable rewards by decomposing complex questions into simpler sub-questions, improving LLM reasoning through guided explorati...

🔹 Publication Date: Published on Jan 31

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.00759
• PDF: https://arxiv.org/pdf/2602.00759

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Rethinking LLM-as-a-Judge: Representation-as-a-Judge with Small Language Models via Semantic Capacity Asymmetry

📝 Summary:
Small language models can effectively evaluate outputs by leveraging internal representations rather than generating responses, enabling a more efficient and interpretable evaluation approach through ...

🔹 Publication Date: Published on Jan 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22588
• PDF: https://arxiv.org/pdf/2601.22588
• Github: https://github.com/zhuochunli/Representation-as-a-judge

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨WildGraphBench: Benchmarking GraphRAG with Wild-Source Corpora

📝 Summary:
WildGraphBench evaluates GraphRAG performance in realistic scenarios using Wikipedia's structured content to assess multi-fact aggregation and summarization capabilities across diverse document types....

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02053
• PDF: https://arxiv.org/pdf/2602.02053

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨RLAnything: Forge Environment, Policy, and Reward Model in Completely Dynamic RL System

📝 Summary:
RLAnything enhances reinforcement learning for LLMs and agents through dynamic model optimization and closed-loop feedback mechanisms that improve policy and reward model training. AI-generated summar...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02488
• PDF: https://arxiv.org/pdf/2602.02488
• Project Page: https://huggingface.co/collections/Gen-Verse/open-agentrl
• Github: https://github.com/Gen-Verse/Open-AgentRL

🔹 Models citing this paper:
• https://huggingface.co/Gen-Verse/RLAnything-Alf-7B
• https://huggingface.co/Gen-Verse/RLAnything-Alf-Reward-14B
• https://huggingface.co/Gen-Verse/RLAnything-OS-Reward-8B

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Wiki Live Challenge: Challenging Deep Research Agents with Expert-Level Wikipedia Articles

📝 Summary:
Deep Research Agents demonstrate capabilities in autonomous information retrieval but show significant gaps when evaluated against expert-level Wikipedia articles using a new live benchmark and compre...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.01590
• PDF: https://arxiv.org/pdf/2602.01590

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research