✨Can Large Language Models Keep Up? Benchmarking Online Adaptation to Continual Knowledge Streams

📝 Summary:
OAKS is a new benchmark to test how LLMs adapt to real-time, evolving information streams. Current models struggle significantly, showing delays and distraction in tracking dynamic knowledge.

🔹 Publication Date: Published on Mar 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.07392
• PDF: https://arxiv.org/pdf/2603.07392
• Github: https://github.com/kaistAI/OAKS

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨CodePercept: Code-Grounded Visual STEM Perception for MLLMs

📝 Summary:
MLLMs struggle with STEM visual reasoning due to perceptual limitations rather than reasoning deficiencies, and enhancing perception through code-as-perception paradigms improves performance. AI-gener...

🔹 Publication Date: Published on Mar 11

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.10757
• PDF: https://arxiv.org/pdf/2603.10757
• Github: https://github.com/TongkunGuan/Qwen-CodePercept

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨MA-EgoQA: Question Answering over Egocentric Videos from Multiple Embodied Agents

📝 Summary:
Multi-agent systems require understanding multiple long-horizon egocentric videos simultaneously, necessitating new benchmarks and models for system-level comprehension. AI-generated summary As embodi...

🔹 Publication Date: Published on Mar 10

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.09827
• PDF: https://arxiv.org/pdf/2603.09827
• Project Page: https://ma-egoqa.github.io/
• Github: https://github.com/KangsanKim07/MA-EgoQA

✨ Datasets citing this paper:
• https://huggingface.co/datasets/KangsanKim71/MA-EgoQA

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Any to Full: Prompting Depth Anything for Depth Completion in One Stage

📝 Summary:
A novel one-stage depth completion framework that uses scale-prompting adaptation of pretrained monocular depth estimation models to handle varying depth sparsity and irregular distributions more effi...

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05711
• PDF: https://arxiv.org/pdf/2603.05711

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨LLM2Vec-Gen: Generative Embeddings from Large Language Models

📝 Summary:
LLM2Vec-Gen introduces a self-supervised method for text embedding that represents model responses through trainable special tokens, achieving superior performance on MTEB while reducing harmful conte...

🔹 Publication Date: Published on Mar 11

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.10913
• PDF: https://arxiv.org/pdf/2603.10913
• Project Page: https://mcgill-nlp.github.io/llm2vec-gen/
• Github: https://github.com/McGill-NLP/llm2vec-gen

🔹 Models citing this paper:
• https://huggingface.co/McGill-NLP/LLM2Vec-Gen-Qwen3-06B
• https://huggingface.co/McGill-NLP/LLM2Vec-Gen-Qwen3-17B
• https://huggingface.co/McGill-NLP/LLM2Vec-Gen-Qwen3-4B

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨TADA: A Generative Framework for Speech Modeling via Text-Acoustic Dual Alignment

📝 Summary:
A novel tokenization scheme synchronizes acoustic features with text tokens in TTS systems, enabling unified modeling and reduced hallucinations through flow matching and text-only guidance. AI-genera...

🔹 Publication Date: Published on Feb 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.23068
• PDF: https://arxiv.org/pdf/2602.23068
• Project Page: https://www.hume.ai/blog/opensource-tada
• Github: https://github.com/HumeAI/tada

🔹 Models citing this paper:
• https://huggingface.co/HumeAI/tada-1b
• https://huggingface.co/HumeAI/tada-3b-ml
• https://huggingface.co/HumeAI/tada-codec

✨ Spaces citing this paper:
• https://huggingface.co/spaces/HumeAI/tada

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Flash-KMeans: Fast and Memory-Efficient Exact K-Means

📝 Summary:
Flash-kmeans enables efficient online k-means clustering on GPUs through novel kernel-level optimizations that eliminate I/O bottlenecks and atomic write contention. AI-generated summary k-means has h...

🔹 Publication Date: Published on Mar 10

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.09229
• PDF: https://arxiv.org/pdf/2603.09229
• Github: https://github.com/svg-project/flash-kmeans

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨ReMix: Reinforcement routing for mixtures of LoRAs in LLM finetuning

📝 Summary:
Researchers address imbalance in routing weights of Mixture-of-LoRAs models by proposing Reinforcement Routing (ReMix), which uses non-learnable weights and reinforcement learning techniques to improv...

🔹 Publication Date: Published on Mar 10

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.10160
• PDF: https://arxiv.org/pdf/2603.10160

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Prism-Δ: Differential Subspace Steering for Prompt Highlighting in Large Language Models

📝 Summary:
PRISM-Δ extracts discriminative steering directions by decomposing cross-covariance differences, uses softplus weights for attention heads, and extends to value representations for improved long-conte...

🔹 Publication Date: Published on Mar 11

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.10705
• PDF: https://arxiv.org/pdf/2603.10705
• Project Page: https://yuyaoge.github.io/PRISM-DELTA/
• Github: https://github.com/YuyaoGe/PRISM-DELTA

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨In-Context Reinforcement Learning for Tool Use in Large Language Models

📝 Summary:
In-Context Reinforcement Learning ICRL is an RL-only framework for LLMs to use external tools, eliminating costly supervised fine-tuning. It teaches tool use through in-context examples during training, gradually reducing them. ICRL proves to be a scalable, data-efficient, and state-of-the-art ap...

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.08068
• PDF: https://arxiv.org/pdf/2603.08068
• Github: https://github.com/applese233/ICRL

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Hindsight Credit Assignment for Long-Horizon LLM Agents

📝 Summary:
HCAPO improves credit assignment in long-horizon LLM agents by using hindsight reasoning to refine Q-values and a multi-scale advantage mechanism. It significantly outperforms state-of-the-art methods, boosting success rates on benchmarks like WebShop and ALFWorld. This enhances exploration and c...

🔹 Publication Date: Published on Mar 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.08754
• PDF: https://arxiv.org/pdf/2603.08754

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

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

#LLMAgents #ReinforcementLearning #AI #MachineLearning #HindsightReasoning

✨UniCom: Unified Multimodal Modeling via Compressed Continuous Semantic Representations

📝 Summary:
UniCom unifies multimodal understanding and generation via compressed continuous semantic representations. It resolves issues with discrete tokenizers and unstable continuous modeling by efficiently reducing channel dimensions. This yields state-of-the-art generation, superior controllability, an...

🔹 Publication Date: Published on Mar 11

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.10702
• PDF: https://arxiv.org/pdf/2603.10702
• Project Page: https://miazhao7708.github.io/UniComPage/

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

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

#MultimodalAI #GenerativeAI #DeepLearning #AIResearch #SemanticRepresentations

✨Lost in Backpropagation: The LM Head is a Gradient Bottleneck

📝 Summary:
The softmax bottleneck in neural LMs is a critical optimization bottleneck, not just an expressivity issue. The rank-D output layer suppresses 95-99% of gradient norm, leading to suboptimal updates and inefficient training. This necessitates new LM head designs.

🔹 Publication Date: Published on Mar 10

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.10145
• PDF: https://arxiv.org/pdf/2603.10145

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

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

#LLM #DeepLearning #Optimization #NeuralNetworks #GradientBottleneck

✨StyleVLA: Driving Style-Aware Vision Language Action Model for Autonomous Driving

📝 Summary:
StyleVLA is a physics-informed VLA model that generates diverse, style-aware, and kinematically plausible driving trajectories. It uses a hybrid loss and a large dataset, outperforming proprietary models like Gemini-3-Pro on specialized driving tasks.

🔹 Publication Date: Published on Mar 10

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.09482
• PDF: https://arxiv.org/pdf/2603.09482

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

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

#AutonomousDriving #VLA #AI #DeepLearning #Robotics

✨Causal Concept Graphs in LLM Latent Space for Stepwise Reasoning

📝 Summary:
Causal Concept Graphs identify causal relationships between concepts in LLMs using sparse autoencoders and differentiable structure learning. This method significantly improves causal fidelity for multi-step reasoning over prior techniques, yielding sparse and stable graphs.

🔹 Publication Date: Published on Mar 11

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.10377
• PDF: https://arxiv.org/pdf/2603.10377

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

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

#CausalAI #LLMs #MachineLearning #GraphLearning #ExplainableAI

✨According to Me: Long-Term Personalized Referential Memory QA

📝 Summary:
ATM-Bench is a new benchmark for multimodal multi-source personalized referential memory QA, addressing limitations of existing dialogue-focused benchmarks. It includes 4 years of personal data and introduces Schema-Guided Memory SGM. Current AI systems perform poorly under 20 percent on hard set...

🔹 Publication Date: Published on Mar 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.01990
• PDF: https://arxiv.org/pdf/2603.01990
• Project Page: https://atmbench.github.io/
• Github: https://github.com/JingbiaoMei/ATM-Bench

✨ Datasets citing this paper:
• https://huggingface.co/datasets/Jingbiao/ATM-Bench

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

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

#AI #QuestionAnswering #LongTermMemory #MachineLearning #Benchmark

✨ID-LoRA: Identity-Driven Audio-Video Personalization with In-Context LoRA

📝 Summary:
ID-LoRA jointly generates visual appearance and voice with a single model, improving personalization. It uses in-context LoRA adaptation and identity guidance to preserve speaker characteristics. This outperforms existing methods in human preference for voice and style similarity.

🔹 Publication Date: Published on Mar 10

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.10256
• PDF: https://arxiv.org/pdf/2603.10256
• Project Page: https://id-lora.github.io/
• Github: https://github.com/ID-LoRA/ID-LoRA

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

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

#GenerativeAI #AudioVisual #LoRA #Personalization #DeepLearning

✨COMIC: Agentic Sketch Comedy Generation

📝 Summary:
An AI system generates comedic videos similar to sketch shows. It employs agent-based optimization and LLM critics, trained on YouTube comedy, to evaluate humor. This system produces content approaching professional sketch quality.

🔹 Publication Date: Published on Mar 11

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.11048
• PDF: https://arxiv.org/pdf/2603.11048
• Project Page: https://susunghong.github.io/COMIC/
• Github: https://susunghong.github.io/COMIC/

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

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

#AI #GenerativeAI #SketchComedy #LLMs #ComputationalCreativity

Follow the Machine Learning with Python channel on WhatsApp: https://whatsapp.com/channel/0029VaC7Weq29753hpcggW2A

✨Meissa: Multi-modal Medical Agentic Intelligence

📝 Summary:
Meissa is a lightweight 4B medical MM-LLM that achieves offline agentic capabilities by distilling trajectories from frontier models. It resolves high cost, latency, and privacy issues, matching or exceeding proprietary agents on 13 medical benchmarks with 25x fewer parameters and 22x lower latency.

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.09018
• PDF: https://arxiv.org/pdf/2603.09018
• Github: https://github.com/Schuture/Meissa

🔹 Models citing this paper:
• https://huggingface.co/CYX1998/Meissa-4B

✨ Datasets citing this paper:
• https://huggingface.co/datasets/CYX1998/Meissa-SFT

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

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

#MedicalAI #LLM #AIagents #MultiModalAI #Healthcare

✨SVG-EAR: Parameter-Free Linear Compensation for Sparse Video Generation via Error-aware Routing

📝 Summary:
SVG-EAR introduces a parameter-free method for video diffusion transformers to reduce quadratic attention cost. It recovers missing contributions via centroid approximation and uses error-aware routing to prioritize high-error blocks. This improves efficiency and quality, achieving significant sp...

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.08982
• PDF: https://arxiv.org/pdf/2603.08982

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

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

#VideoGeneration #DiffusionModels #Transformers #AIResearch #MachineLearning