✨Instruction Anchors: Dissecting the Causal Dynamics of Modality Arbitration

📝 Summary:
Instruction tokens act as anchors for modality arbitration in MLLMs, guiding multimodal context use. This involves shallow layers gathering cues and deep layers resolving competition. Manipulating a few specialized attention heads significantly impacts this process.

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.03677
• PDF: https://arxiv.org/pdf/2602.03677

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

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

#MLLMs #MultimodalAI #AttentionMechanisms #DeepLearning #AIResearch

✨RecGOAT: Graph Optimal Adaptive Transport for LLM-Enhanced Multimodal Recommendation with Dual Semantic Alignment

📝 Summary:
RecGOAT bridges the representational gap between LLMs and recommendation systems. It uses graph attention networks and a dual-granularity semantic alignment framework combining cross-modal contrastive learning and optimal adaptive transport for superior performance.

🔹 Publication Date: Published on Jan 31

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.00682
• PDF: https://arxiv.org/pdf/2602.00682
• Github: https://github.com/6lyc/RecGOAT-LLM4Rec

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

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

#RecGOAT #LLM #RecommendationSystems #MultimodalAI #GraphNeuralNetworks

✨POP: Prefill-Only Pruning for Efficient Large Model Inference

📝 Summary:
POP is a new stage-aware pruning method for large models. It omits deep layers during the computationally intensive prefill stage while using the full model for decoding. This achieves up to 1.37 times prefill speedup with minimal accuracy loss, overcoming limitations of prior pruning methods.

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.03295
• PDF: https://arxiv.org/pdf/2602.03295

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

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

#AI #MachineLearning #LLM #ModelPruning #InferenceOptimization

✨MEG-XL: Data-Efficient Brain-to-Text via Long-Context Pre-Training

📝 Summary:
MEG-XL improves brain-to-text decoding by pre-training with 2.5 minutes of MEG context, far exceeding prior methods. This long-context approach dramatically boosts data efficiency, achieving supervised performance with only a fraction of the data and outperforming other models.

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02494
• PDF: https://arxiv.org/pdf/2602.02494
• Github: https://github.com/neural-processing-lab/MEG-XL

🔹 Models citing this paper:
• https://huggingface.co/pnpl/MEG-XL

✨ Datasets citing this paper:
• https://huggingface.co/datasets/pnpl/LibriBrain

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

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

#BrainToText #MEG #Neuroscience #DeepLearning #AI

✨LangMap: A Hierarchical Benchmark for Open-Vocabulary Goal Navigation

📝 Summary:
HieraNav introduces a multi-granularity, open-vocabulary navigation task. LangMap, its benchmark, uses 3D scans and human annotations across four semantic levels. Evaluations highlight challenges for models in complex navigation goals.

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02220
• PDF: https://arxiv.org/pdf/2602.02220
• Project Page: https://bo-miao.github.io/LangMap/
• Github: https://github.com/bo-miao/LangMap

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

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

#AINavigation #ComputerVision #Robotics #NLP #Benchmark

✨MedSAM-Agent: Empowering Interactive Medical Image Segmentation with Multi-turn Agentic Reinforcement Learning

📝 Summary:
MedSAM-Agent reformulates medical image segmentation as a multi-step decision-making process using hybrid prompting and a two-stage training pipeline with process rewards to improve autonomous reasoni...

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.03320
• PDF: https://arxiv.org/pdf/2602.03320
• Github: https://github.com/CUHK-AIM-Group/MedSAM-Agent

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨You Need an Encoder for Native Position-Independent Caching

📝 Summary:
LLM KV caches are inefficient for arbitrary context orders. This paper proposes native PIC by reintroducing an encoder to decoder-only LLMs and developing COMB a PIC-aware caching system. COMB reduces TTFT by 51-94 percent and triples throughput with comparable accuracy.

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.01519
• PDF: https://arxiv.org/pdf/2602.01519

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

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

#LLM #Caching #DeepLearning #AI #Performance

✨Neural Predictor-Corrector: Solving Homotopy Problems with Reinforcement Learning

📝 Summary:
Neural Predictor-Corrector NPC unifies diverse homotopy problems, using reinforcement learning to learn optimal policies. This general neural solver consistently outperforms classical methods in efficiency and stability across tasks.

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.03086
• PDF: https://arxiv.org/pdf/2602.03086

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

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

#ReinforcementLearning #HomotopyProblems #NeuralNetworks #MachineLearning #AI

✨RANKVIDEO: Reasoning Reranking for Text-to-Video Retrieval

📝 Summary:
RANKVIDEO is a reasoning-based reranker for text-to-video retrieval that explicitly analyzes query-video pairs for relevance. It uses a multi-objective training approach and a data synthesis pipeline. RANKVIDEO significantly improves retrieval performance by 31 percent on a large benchmark, outpe...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02444
• PDF: https://arxiv.org/pdf/2602.02444
• Github: https://github.com/tskow99/RANKVIDEO-Reasoning-Reranker

🔹 Models citing this paper:
• https://huggingface.co/hltcoe/RankVideo

✨ Datasets citing this paper:
• https://huggingface.co/datasets/hltcoe/RankVideo-Dataset

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨LIVE: Long-horizon Interactive Video World Modeling

📝 Summary:
LIVE is a long-horizon video world model that uses cycle-consistency and diffusion loss to control error accumulation during extended video generation. AI-generated summary Autoregressive video world ...

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/pdf/2602.03747
• PDF: https://arxiv.org/pdf/2602.03747
• Project Page: https://junchao-cs.github.io/LIVE-demo/
• Github: https://junchao-cs.github.io/LIVE-demo/

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

🎯 Want to Upskill in IT? Try Our FREE 2026 Learning Kits!

SPOTO gives you free, instant access to high-quality, updated resources that help you study smarter and pass exams faster.
✅ Latest Exam Materials:
Covering #Python, #Cisco, #PMI, #Fortinet, #AWS, #Azure, #AI, #Excel, #comptia, #ITIL, #cloud & more!
✅ 100% Free, No Sign-up:
All materials are instantly downloadable

✅ What’s Inside:
・📘IT Certs E-book: https://bit.ly/3Mlu5ez
・📝IT Exams Skill Test: https://bit.ly/3NVrgRU
・🎓Free IT courses: https://bit.ly/3M9h5su
・🤖Free PMP Study Guide: https://bit.ly/4te3EIn
・☁️Free Cloud Study Guide: https://bit.ly/4kgFVDs

👉 Become Part of Our IT Learning Circle! resources and support:
https://chat.whatsapp.com/FlG2rOYVySLEHLKXF3nKGB

💬 Want exam help? Chat with an admin now!
wa.link/8fy3x4

✨Didactic to Constructive: Turning Expert Solutions into Learnable Reasoning

📝 Summary:
DAIL improves LLM reasoning by converting didactic expert solutions into detailed, in-distribution traces via contrastive learning. This method achieves 10-25% performance gains and 2-4x reasoning efficiency using minimal expert data.

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02405
• PDF: https://arxiv.org/pdf/2602.02405
• Github: https://github.com/ethanm88/DAIL

✨ Datasets citing this paper:
• https://huggingface.co/datasets/emendes3/e1-proof

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Feedback by Design: Understanding and Overcoming User Feedback Barriers in Conversational Agents

📝 Summary:
High-quality feedback is essential for effective human-AI interaction. It bridges knowledge gaps, corrects digressions, and shapes system behavior; both during interaction and throughout model develop...

🔹 Publication Date: Published on Feb 1

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.01405
• PDF: https://arxiv.org/pdf/2602.01405

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Scaling Small Agents Through Strategy Auctions

📝 Summary:
Small language models fail on complex tasks. The paper proposes Strategy Auctions for Workload Efficiency SALE, a marketplace-inspired framework where agents bid strategic plans for task routing and self-improvement. SALE reduces costs by 35% and improves performance, enabling small agents to sca...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02751
• PDF: https://arxiv.org/pdf/2602.02751

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨MemoryLLM: Plug-n-Play Interpretable Feed-Forward Memory for Transformers

📝 Summary:
MemoryLLM decouples feed-forward networks from self-attention in transformers, enabling context-free token-wise neural retrieval memory that improves inference efficiency through pre-computed lookups....

🔹 Publication Date: Published on Jan 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.00398
• PDF: https://arxiv.org/pdf/2602.00398

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Position: Agentic Evolution is the Path to Evolving LLMs

📝 Summary:
Large language models struggle to adapt to changing real-world environments. Agentic evolution is proposed as a new approach where deployment-time improvement becomes a goal-directed optimization process. This allows for sustained, open-ended adaptation by scaling evolution.

🔹 Publication Date: Published on Jan 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.00359
• PDF: https://arxiv.org/pdf/2602.00359
• Github: https://github.com/ventr1c/agentic-evoluiton

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe

📝 Summary:
MiniCPM-V 4.5 is an 8B multimodal LLM achieving high performance and efficiency. It uses a unified 3D-Resampler, unified learning, and hybrid reinforcement learning. It surpasses larger models like GPT-4o and Qwen2.5-VL with significantly less memory and faster inference.

🔹 Publication Date: Published on Sep 16, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2509.18154
• PDF: https://arxiv.org/pdf/2509.18154
• Github: https://github.com/OpenBMB/MiniCPM-V

🔹 Models citing this paper:
• https://huggingface.co/openbmb/MiniCPM-V-4_5
• https://huggingface.co/openbmb/MiniCPM-V-4_5-gguf
• https://huggingface.co/openbmb/MiniCPM-V-4_5-AWQ

✨ Datasets citing this paper:
• https://huggingface.co/datasets/openbmb/RLAIF-V-Dataset

✨ Spaces citing this paper:
• https://huggingface.co/spaces/CGQN/MiniCPM-V-4_5-int4-CPU-0
• https://huggingface.co/spaces/CGQN/MiniCPM-V-4_5-from_gpt5
• https://huggingface.co/spaces/CGQN/MiniCPM-V-4_5

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Privasis: Synthesizing the Largest "Public" Private Dataset from Scratch

📝 Summary:
Privasis is a new million-scale synthetic dataset for AI privacy research. It addresses data scarcity, enabling compact sanitization models that outperform large language models like GPT-5. The diverse dataset and models will be released to the public.

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.03183
• PDF: https://arxiv.org/pdf/2602.03183
• Project Page: https://privasis.github.io
• Github: https://github.com/skywalker023/privasis

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨FIRE-Bench: Evaluating Agents on the Rediscovery of Scientific Insights

📝 Summary:
FIRE-Bench evaluates AI agents on rediscovering scientific findings through full research cycles, from hypothesis to conclusions. Agents receive a high-level question and act autonomously. Current agents struggle, showing that reliable AI-driven scientific discovery remains challenging.

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02905
• PDF: https://arxiv.org/pdf/2602.02905
• Project Page: https://firebench.github.io/

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨UI-TARS-2 Technical Report: Advancing GUI Agent with Multi-Turn Reinforcement Learning

📝 Summary:
UI-TARS-2 is a native GUI agent model that tackles challenges in data scalability and multi-turn reinforcement learning. It significantly improves over its predecessor and strong baselines on GUI and game benchmarks, demonstrating robust generalization. This advances GUI agents for real-world int...

🔹 Publication Date: Published on Sep 2, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2509.02544
• PDF: https://arxiv.org/pdf/2509.02544
• Project Page: https://seed-tars.com/showcase/ui-tars-2/
• Github: https://github.com/bytedance/ui-tars

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research