✨GEBench: Benchmarking Image Generation Models as GUI Environments

📝 Summary:
This paper introduces GEBench, a new benchmark and GE-Score metric for evaluating temporal coherence and dynamic interaction in GUI generation models. Evaluations show current models struggle significantly with consistency and grounding over longer interaction sequences.

🔹 Publication Date: Published on Feb 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.09007
• PDF: https://arxiv.org/pdf/2602.09007
• Github: https://github.com/stepfun-ai/GEBench

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

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

#ImageGeneration #GUIGeneration #AIResearch #Benchmarking #MachineLearning

✨Thinking Makes LLM Agents Introverted: How Mandatory Thinking Can Backfire in User-Engaged Agents

📝 Summary:
Mandatory explicit thinking in user-engaged LLM agents often degrades performance. This occurs because thinking makes agents introverted, shortening responses and reducing information disclosure. Prompting for transparency significantly improves agent performance by enhancing communication.

🔹 Publication Date: Published on Feb 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.07796
• PDF: https://arxiv.org/pdf/2602.07796

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

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

#LLMAgents #AIResearch #PromptEngineering #HumanAIInteraction #AIBehavior

✨FlexMoRE: A Flexible Mixture of Rank-heterogeneous Experts for Efficient Federatedly-trained Large Language Models

📝 Summary:
FlexMoRE proposes replacing full-sized experts with low-rank adapters in Mixture-of-Experts for federated LLMs. This flexible approach improves performance using significantly fewer parameters, with optimal expert rank depending on task complexity.

🔹 Publication Date: Published on Feb 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.08818
• PDF: https://arxiv.org/pdf/2602.08818

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

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

#LLM #FederatedLearning #MixtureOfExperts #AI #DeepLearning

✨GraphAgents: Knowledge Graph-Guided Agentic AI for Cross-Domain Materials Design

📝 Summary:
GraphAgents is a multi-agent AI framework using knowledge graphs to solve complex materials design problems. It deploys specialized agents for tasks like evidence retrieval and graph traversal, outperforming single-shot LLMs. This approach effectively identifies sustainable PFAS alternatives, exp...

🔹 Publication Date: Published on Feb 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.07491
• PDF: https://arxiv.org/pdf/2602.07491

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

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

#AI #KnowledgeGraphs #AgenticAI #MaterialsDesign #MultiAgentSystems

✨On Randomness in Agentic Evals

📝 Summary:
Agentic system evaluations using single-run pass@1 scores are highly unreliable due to significant variance, often masking genuine progress. Small reported improvements may reflect evaluation noise. Reliable assessment requires multiple runs, statistical analysis, and metrics like pass@k.

🔹 Publication Date: Published on Feb 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.07150
• PDF: https://arxiv.org/pdf/2602.07150

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

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

#AIEvaluation #AgenticAI #MachineLearning #StatisticalMethods #AIResearch

✨Echoes as Anchors: Probabilistic Costs and Attention Refocusing in LLM Reasoning

📝 Summary:
This paper formalizes the Echo of Prompt EOP, spontaneous question repetition by LLMs, as a compute-shaping mechanism. It introduces Echo-Distilled SFT and Echoic Prompting to leverage EOP, improving reasoning accuracy and efficiency by refocusing attention.

🔹 Publication Date: Published on Feb 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.06600
• PDF: https://arxiv.org/pdf/2602.06600
• Github: https://github.com/hhh2210/echoes-as-anchors

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

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

#LLM #PromptEngineering #AIResearch #DeepLearning #AIAttention

✨AIRS-Bench: a Suite of Tasks for Frontier AI Research Science Agents

📝 Summary:
AIRS-Bench is a new benchmark of 20 scientific tasks evaluating AI agents across the full research lifecycle. Agents exceed human state-of-the-art in 4 tasks but largely fall short, highlighting significant room for improvement in autonomous scientific research. The suite is open-sourced to accel...

🔹 Publication Date: Published on Feb 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.06855
• PDF: https://arxiv.org/pdf/2602.06855
• Github: https://github.com/facebookresearch/airs-bench

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

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

#AIagents #ScientificResearch #AIBenchmark #FrontierAI #AutonomousResearch

✨Fundamental Reasoning Paradigms Induce Out-of-Domain Generalization in Language Models

📝 Summary:
This study explores how fundamental reasoning paradigms deduction induction and abduction influence LLM generalization. By training LLMs on a new dataset of symbolic reasoning trajectories, the research shows substantial performance gains and strong generalizability on realistic out-of-domain tasks.

🔹 Publication Date: Published on Feb 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.08658
• PDF: https://arxiv.org/pdf/2602.08658
• Github: https://github.com/voalmciaf/FR-OOD

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

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

#LLM #AI #MachineLearning #Reasoning #Generalization

✨Data Science and Technology Towards AGI Part I: Tiered Data Management

📝 Summary:
This paper proposes an LLM-guided, tiered data management framework L0-L4 to optimize data quality, acquisition cost, and training efficiency. This systematic approach, used across LLM development stages, significantly improves model performance and sustainability.

🔹 Publication Date: Published on Feb 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.09003
• PDF: https://arxiv.org/pdf/2602.09003
• Project Page: https://ultradata.openbmb.cn/
• Github: https://github.com/UltraData-OpenBMB/UltraData-Math

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

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

#DataScience #LLM #AGI #DataManagement #AIResearch

✨Context Compression via Explicit Information Transmission

📝 Summary:
ComprExIT enhances LLM long-context inference via explicit information transmission over frozen hidden states. This lightweight method uses depth-wise and width-wise transmission to mitigate overwriting and coordinate information allocation, outperforming existing compression techniques with mini...

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.03784
• PDF: https://arxiv.org/pdf/2602.03784

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨dewi-kadita: A Python Library for Idealized Fish Schooling Simulation with Entropy-Based Diagnostics

📝 Summary:
Collective motion in fish schools exemplifies emergent self-organization in active matter systems, yet computational tools for simulating and analyzing these dynamics remain fragmented across research...

🔹 Publication Date: Published on Feb 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.07948
• PDF: https://arxiv.org/pdf/2602.07948
• Project Page: https://pypi.org/project/dewi-kadita/
• Github: https://github.com/sandyherho/dewi-kadita

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Cost-Efficient RAG for Entity Matching with LLMs: A Blocking-based Exploration

📝 Summary:
CE-RAG4EM reduces computational overhead in large-scale entity matching by implementing blocking-based batch retrieval and generation while maintaining competitive matching quality. AI-generated summa...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05708
• PDF: https://arxiv.org/pdf/2602.05708

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Statistical Learning Theory in Lean 4: Empirical Processes from Scratch

📝 Summary:
A comprehensive formalization of statistical learning theory in Lean 4 addresses gaps in mathematical libraries and demonstrates human-AI collaboration for verified machine learning theory foundations...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02285
• PDF: https://arxiv.org/pdf/2602.02285
• Github: https://github.com/YuanheZ/lean-stat-learning-theory

✨ Datasets citing this paper:
• https://huggingface.co/datasets/liminho123/lean4-stat-learning-theory-novel
• https://huggingface.co/datasets/liminho123/lean4-stat-learning-theory-random
• https://huggingface.co/datasets/liminho123/lean4-stat-learning-theory-corpus

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Optimal Turkish Subword Strategies at Scale: Systematic Evaluation of Data, Vocabulary, Morphology Interplay

📝 Summary:
This study systematically evaluates Turkish subword tokenization, varying vocabulary and corpus size across multiple tokenizer families and diverse linguistic tasks. It introduces morphology-aware diagnostics to provide actionable guidance for building effective tokenizers in morphologically rich...

🔹 Publication Date: Published on Feb 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.06942
• PDF: https://arxiv.org/pdf/2602.06942

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨KV-CoRE: Benchmarking Data-Dependent Low-Rank Compressibility of KV-Caches in LLMs

📝 Summary:
KV-CoRE method evaluates kv-cache compressibility through SVD-based low-rank approximation, revealing patterns linking compressibility to model architecture and training data across multiple languages...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05929
• PDF: https://arxiv.org/pdf/2602.05929

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨CauScale: Neural Causal Discovery at Scale

📝 Summary:
CauScale is a neural architecture that enables efficient causal discovery on large graphs through compressed embeddings and tied attention weights, achieving high accuracy and significant speedups ove...

🔹 Publication Date: Published on Feb 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.08629
• PDF: https://arxiv.org/pdf/2602.08629

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨f-GRPO and Beyond: Divergence-Based Reinforcement Learning Algorithms for General LLM Alignment

📝 Summary:
Preference alignment objectives are extended to general alignment settings using f-divergence variational representations, introducing novel on-policy and hybrid policy optimization methods for LLM al...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05946
• PDF: https://arxiv.org/pdf/2602.05946

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Agent Skills: A Data-Driven Analysis of Claude Skills for Extending Large Language Model Functionality

📝 Summary:
Agent skills extend large language model (LLM) agents with reusable, program-like modules that define triggering conditions, procedural logic, and tool interactions. As these skills proliferate in pub...

🔹 Publication Date: Published on Feb 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.08004
• PDF: https://arxiv.org/pdf/2602.08004

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨CodeCircuit: Toward Inferring LLM-Generated Code Correctness via Attribution Graphs

📝 Summary:
CodeCircuit assesses LLM code correctness purely from its internal neural dynamics. It uses algorithmic attribution graphs to identify structural signatures of correct reasoning versus failure. This reliably predicts correctness and fixes errors.

🔹 Publication Date: Published on Feb 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.07080
• PDF: https://arxiv.org/pdf/2602.07080
• Github: https://github.com/bruno686/CodeCircuit

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Towards Agentic Intelligence for Materials Science

📝 Summary:
AI-driven materials science integrates large language models across discovery pipelines from data curation to agent-based experimentation, emphasizing system-level optimization and autonomous goal pur...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.00169
• PDF: https://arxiv.org/pdf/2602.00169

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Anchored Decoding: Provably Reducing Copyright Risk for Any Language Model

📝 Summary:
Anchored Decoding is an inference-time method that reduces verbatim copying in language models. It guides a risky LM with a permissively trained safe LM, significantly lowering copyright risk while preserving fluency and factuality. This method achieves up to 75% reduction in measurable copying.

🔹 Publication Date: Published on Feb 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.07120
• PDF: https://arxiv.org/pdf/2602.07120
• Project Page: https://tinyurl.com/anchored-decoding-demo
• Github: https://github.com/jacqueline-he/anchored-decoding

🔹 Models citing this paper:
• https://huggingface.co/jacquelinehe/tinycomma-1.8b-llama3-tokenizer

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

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

#LLM #AICopyright #AISafety #ResponsibleAI #AIResearch