ML Research Hub
32.8K subscribers
5.54K photos
351 videos
24 files
5.99K links
Advancing research in Machine Learning – practical insights, tools, and techniques for researchers.

Admin: @HusseinSheikho || @Hussein_Sheikho
Download Telegram
CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation

📝 Summary:
CUDA Agent is a large-scale agentic reinforcement learning system for optimizing CUDA kernels. It uses data synthesis, a skill-augmented development environment, and RL to achieve state-of-the-art performance, outperforming torch.compile and other LLMs significantly.

🔹 Publication Date: Published on Feb 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.24286
• PDF: https://arxiv.org/pdf/2602.24286
• Project Page: https://cuda-agent.github.io/

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

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

#CUDA #ReinforcementLearning #HPC #AI #MachineLearning
Memory Caching: RNNs with Growing Memory

📝 Summary:
Memory Caching MC enhances recurrent neural networks by caching memory states, allowing their capacity to grow with sequence length. This bridges the performance gap between RNNs and Transformers in long-context and recall-intensive tasks, outperforming other recurrent models.

🔹 Publication Date: Published on Feb 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.24281
• PDF: https://arxiv.org/pdf/2602.24281

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

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

#RNNs #Transformers #NeuralNetworks #MemoryCaching #AI
Ref-Adv: Exploring MLLM Visual Reasoning in Referring Expression Tasks

📝 Summary:
Ref-Adv is a challenging benchmark for referring expression comprehension that eliminates shortcut solutions by using complex linguistic expressions with minimal identifying information and hard distr...

🔹 Publication Date: Published on Feb 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.23898
• PDF: https://arxiv.org/pdf/2602.23898
• Project Page: https://ref-adv.github.io
• Github: https://github.com/dddraxxx/Ref-Adv

Datasets citing this paper:
https://huggingface.co/datasets/dddraxxx/ref-adv-s

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
DLEBench: Evaluating Small-scale Object Editing Ability for Instruction-based Image Editing Model

📝 Summary:
A new benchmark called DeepLookEditBench is introduced to evaluate instruction-based image editing models' capability in handling small-scale object editing, revealing significant performance gaps in ...

🔹 Publication Date: Published on Feb 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.23622
• PDF: https://arxiv.org/pdf/2602.23622

Datasets citing this paper:
https://huggingface.co/datasets/SPUH/DLEBench

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
dLLM: Simple Diffusion Language Modeling

📝 Summary:
dLLM is an open-source framework standardizing core components of diffusion language modeling. It addresses the issue of scattered, hard-to-reproduce DLM implementations, enabling easy reproduction, customization, and development of both small and large diffusion language models.

🔹 Publication Date: Published on Feb 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.22661
• PDF: https://arxiv.org/pdf/2602.22661
• Project Page: https://github.com/ZHZisZZ/dllm
• Github: https://github.com/ZHZisZZ/dllm

🔹 Models citing this paper:
https://huggingface.co/dllm-hub/ModernBERT-large-chat-v0.1
https://huggingface.co/dllm-hub/Qwen3-0.6B-diffusion-mdlm-v0.1
https://huggingface.co/dllm-hub/Qwen3-0.6B-diffusion-bd3lm-v0.1

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

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

#DiffusionModels #LanguageModeling #LLMs #OpenSourceAI #AIResearch
CiteAudit: You Cited It, But Did You Read It? A Benchmark for Verifying Scientific References in the LLM Era

📝 Summary:
CiteAudit presents a benchmark and multi-agent pipeline to detect fabricated citations, a risk introduced by LLMs in scientific writing. This framework decomposes citation verification into steps and significantly outperforms prior methods, offering scalable tools for trustworthy references.

🔹 Publication Date: Published on Feb 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.23452
• PDF: https://arxiv.org/pdf/2602.23452
• Project Page: https://www.checkcitation.com/

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

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

#LLM #AcademicIntegrity #CitationVerification #AIethics #ResearchTools
Media is too big
VIEW IN TELEGRAM
Mode Seeking meets Mean Seeking for Fast Long Video Generation

📝 Summary:
This paper introduces a Decoupled Diffusion Transformer combining mode seeking and mean seeking for efficient long video generation. It leverages global flow matching for narrative coherence and local distribution matching against a short-video teacher for realism, effectively bridging the fideli...

🔹 Publication Date: Published on Feb 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.24289
• PDF: https://arxiv.org/pdf/2602.24289
• Project Page: https://primecai.github.io/mmm/
• Github: https://primecai.github.io/mmm/

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

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

#VideoGeneration #DiffusionModels #AIResearch #MachineLearning #ComputerVision
LMCache: An Efficient KV Cache Layer for Enterprise-Scale LLM Inference

📝 Summary:
LMCACHE enables efficient KV cache management for large language models by storing caches outside GPU memory, supporting cache reuse across queries and inference engines while achieving significant th...

🔹 Publication Date: Published on Oct 8, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.09665
• PDF: https://arxiv.org/pdf/2510.09665
• Github: https://github.com/LMCache/LMCache

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Recovered in Translation: Efficient Pipeline for Automated Translation of Benchmarks and Datasets

📝 Summary:
This paper introduces an automated framework for high-quality multilingual translation of benchmarks. It uses test-time compute scaling, specifically Universal Self-Improvement and T-RANK, to prevent semantic drift and context loss. This improves LLM evaluation accuracy beyond existing methods.

🔹 Publication Date: Published on Feb 25

🔹 Paper Links:
• arXiv Page: https://huggingface.co/collections/hannayukhymenko/recovered-in-translation-eacl26-mme
• PDF: https://arxiv.org/pdf/2602.22207
• Project Page: https://ritranslation.insait.ai/
• Github: https://github.com/insait-institute/ritranslation

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

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

#LLMEvaluation #MachineTranslation #NLP #AIResearch #Benchmarks
InfoNCE Induces Gaussian Distribution

📝 Summary:
InfoNCE objective induces Gaussian distribution in contrastive learning representations. This is supported by theoretical analysis and experimental validation. It explains observed Gaussianity and enables analytical treatment of representations.

🔹 Publication Date: Published on Feb 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.24012
• PDF: https://arxiv.org/pdf/2602.24012
• Project Page: https://rbetser.github.io/InfoNCE-induces-Gaussian-distribution/
• Github: https://github.com/rbetser/InfoNCE-induces-Gaussian-distribution

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
LongVideo-R1: Smart Navigation for Low-cost Long Video Understanding

📝 Summary:
LongVideo-R1 is an MLLM agent for efficient long video understanding with low computational cost. It uses active reasoning and selective clip navigation, avoiding exhaustive search by focusing on informative segments. This achieves superior accuracy and efficiency.

🔹 Publication Date: Published on Feb 24

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.20913
• PDF: https://arxiv.org/pdf/2602.20913
• Github: https://github.com/qiujihao19/LongVideo-R1

🔹 Models citing this paper:
https://huggingface.co/ChurchillQAQ/LongVideo-R1-Qwen2.5
https://huggingface.co/ChurchillQAQ/LongVideo-R1-Qwen3

Datasets citing this paper:
https://huggingface.co/datasets/ChurchillQAQ/LongVideo-R1-Data

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

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

#VideoUnderstanding #MLLM #AI #EfficientAI #DeepLearning
LK Losses: Direct Acceptance Rate Optimization for Speculative Decoding

📝 Summary:
LK losses directly optimize speculative decoding acceptance rate, outperforming standard KL divergence training. This boosts speedup, showing consistent gains of up to 10% in average acceptance length across various models and domains with no extra overhead.

🔹 Publication Date: Published on Feb 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.23881
• PDF: https://arxiv.org/pdf/2602.23881

🔹 Models citing this paper:
https://huggingface.co/nebius/MTP-DeepSeek-V3-0324
https://huggingface.co/nebius/MEDUSA-Llama-3.1-8B-Instruct
https://huggingface.co/nebius/MLP-Speculator-Llama-3.1-8B-Instruct

Datasets citing this paper:
https://huggingface.co/datasets/nebius/Llama-3.1-8B-Instruct-Infinity-Instruct-0625
https://huggingface.co/datasets/nebius/gpt-oss-20b-Infinity-Instruct-0625
https://huggingface.co/datasets/nebius/DeepSeek-V3-Infinity-Instruct-0625

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

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

#SpeculativeDecoding #LLMs #LLMOptimization #DeepLearning #AIResearch
Compositional Generalization Requires Linear, Orthogonal Representations in Vision Embedding Models

📝 Summary:
Compositional generalization requires neural representations to decompose linearly into orthogonal per-concept components. This Linear Representation Hypothesis is theoretically grounded and empirically supported in vision models.

🔹 Publication Date: Published on Feb 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.24264
• PDF: https://arxiv.org/pdf/2602.24264
• Github: https://github.com/oshapio/necessary-compositionality

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

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

#CompositionalGeneralization #VisionModels #NeuralNetworks #MachineLearning #AIResearch
CL4SE: A Context Learning Benchmark For Software Engineering Tasks

📝 Summary:
CL4SE presents a benchmark for evaluating context learning in software engineering tasks, defining four SE-specific context types. It demonstrates an average 24.7% performance improvement for LLMs across tasks like code generation and review, establishing a standardized evaluation framework.

🔹 Publication Date: Published on Feb 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.23047
• PDF: https://arxiv.org/pdf/2602.23047
• Project Page: https://huggingface.co/papers?q=project-specific%20context
• Github: https://github.com/Tomsawyerhu/CodeCL

Datasets citing this paper:
https://huggingface.co/datasets/tomhu/codecl

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

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

#ContextLearning #SoftwareEngineering #LLMs #CodeGeneration #Benchmarks
1
Fara-7B: An Efficient Agentic Model for Computer Use

📝 Summary:
FaraGen synthesizes high-quality datasets for computer use agents to solve web tasks. This data trains Fara-7B, an efficient model perceiving via screenshots that outperforms larger models on benchmarks. It shows scalable data generation advances small agentic models.

🔹 Publication Date: Published on Nov 24, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2511.19663
• PDF: https://arxiv.org/pdf/2511.19663
• Project Page: https://aka.ms/msaif/fara
• Github: https://github.com/microsoft/fara

🔹 Models citing this paper:
https://huggingface.co/microsoft/Fara-7B
https://huggingface.co/XythicK/microsoft_Fara-7B-GGUF

Datasets citing this paper:
https://huggingface.co/datasets/microsoft/WebTailBench

Spaces citing this paper:
https://huggingface.co/spaces/2025-ai-timeline/2025-ai-timeline
https://huggingface.co/spaces/prithivMLmods/CUA-GUI-Operator
https://huggingface.co/spaces/gouyongxiang/fara-7b

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

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

#AIAgents #MachineLearning #EfficientAI #DatasetGeneration #WebAutomation
How to Take a Memorable Picture? Empowering Users with Actionable Feedback

📝 Summary:
This paper introduces Memorability Feedback MemFeed, a new task providing actionable natural language guidance to improve photo memorability. Their method, MemCoach, uses MLLMs and a teacher-student strategy, demonstrating that memorability can be taught and instructed.

🔹 Publication Date: Published on Feb 25

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.21877
• PDF: https://arxiv.org/pdf/2602.21877
• Project Page: https://laitifranz.github.io/MemCoach/
• Github: https://laitifranz.github.io/MemCoach/

Datasets citing this paper:
https://huggingface.co/datasets/laitifranz/MemBench-InternVL3.5-Eval
https://huggingface.co/datasets/laitifranz/MemBench

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

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

#PhotoMemorability #MLLMs #ComputerVision #AIResearch #HumanComputerInteraction
Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution

📝 Summary:
ReMe is a dynamic memory framework for LLM agents that distills, reuses, and refines experiences. It boosts performance, allowing smaller models to outperform larger memoryless ones for efficient lifelong learning.

🔹 Publication Date: Published on Dec 11, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.10696
• PDF: https://arxiv.org/pdf/2512.10696
• Project Page: https://reme.agentscope.io/
• Github: https://github.com/agentscope-ai/ReMe

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

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

#LLMAgents #LifelongLearning #LLMMemory #ArtificialIntelligence #MachineLearning
DUET-VLM: Dual stage Unified Efficient Token reduction for VLM Training and Inference

📝 Summary:
DUET-VLM proposes a dual-stage compression framework for Vision-Language Models. It first reduces visual tokens from the vision encoder, then progressively drops less informative tokens in the language backbone, guided by text. This maintains high accuracy while significantly reducing computation...

🔹 Publication Date: Published on Feb 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.18846
• PDF: https://arxiv.org/pdf/2602.18846
• Github: https://github.com/AMD-AGI/DUET-VLM

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

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

#VLM #ModelCompression #AI #DeepLearning #Efficiency
Shared Nature, Unique Nurture: PRISM for Pluralistic Reasoning via In-context Structure Modeling

📝 Summary:
LLMs are converging towards a singular 'hivemind,' reducing diversity. PRISM addresses this by equipping models with individualized epistemic trajectories using dynamic on-the-fly epistemic graphs. This enhances creativity, expands diversity, and improves diagnostic accuracy, moving towards a plu...

🔹 Publication Date: Published on Feb 24

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.21317
• PDF: https://arxiv.org/pdf/2602.21317
• Project Page: https://www.prism4research.com/

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

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

#LLMs #ArtificialIntelligence #AIDiversity #EpistemicModeling #AIResearch