ML Research Hub – Telegram
ML Research Hub
32.9K subscribers
4.64K photos
286 videos
24 files
5.01K links
Advancing research in Machine Learning – practical insights, tools, and techniques for researchers.

Admin: @HusseinSheikho || @Hussein_Sheikho
Download Telegram
TwinBrainVLA: Unleashing the Potential of Generalist VLMs for Embodied Tasks via Asymmetric Mixture-of-Transformers

📝 Summary:
TwinBrainVLA resolves the VLM tension in robot control by coordinating a frozen generalist VLM Left Brain with a trainable specialist VLM Right Brain via Asymmetric Mixture-of-Transformers. This approach achieves superior manipulation performance while preserving semantic understanding for genera...

🔹 Publication Date: Published on Jan 20

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.14133
• PDF: https://arxiv.org/pdf/2601.14133
• Github: https://github.com/ZGC-EmbodyAI/TwinBrainVLA

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#VLM #EmbodiedAI #Robotics #Transformers #AIResearch
Media is too big
VIEW IN TELEGRAM
VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents

📝 Summary:
VisGym introduces 17 environments to evaluate VLM performance in multi-step visual interactions. Current models struggle, especially with long contexts and visual symbolic tasks. Explicit goals and demonstrations offer pathways for improvement.

🔹 Publication Date: Published on Jan 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.16973
• PDF: https://arxiv.org/pdf/2601.16973
• Project Page: https://visgym.github.io/
• Github: https://visgym.github.io/

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#MultimodalAI #VisualLanguageModels #AIenvironments #ComputerVision #AIResearch
LongCat-Flash-Thinking-2601 Technical Report

📝 Summary:
LongCat-Flash-Thinking-2601 is a 560B MoE reasoning model that achieves state-of-the-art performance on agentic benchmarks. Its capabilities stem from a unified training framework, robust tool interaction, and a Heavy Thinking mode for complex reasoning.

🔹 Publication Date: Published on Jan 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.16725
• PDF: https://arxiv.org/pdf/2601.16725

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#MoE #ReasoningModels #AgentAI #LLM #AI
Endless Terminals: Scaling RL Environments for Terminal Agents

📝 Summary:
Endless Terminals introduces an autonomous pipeline for generating procedural terminal tasks that significantly improves agent performance on both synthetic and human-curated benchmarks through scalab...

🔹 Publication Date: Published on Jan 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.16443
• PDF: https://arxiv.org/pdf/2601.16443
• Github: https://github.com/kanishkg/endless-terminals

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research
DSGym: A Holistic Framework for Evaluating and Training Data Science Agents

📝 Summary:
DSGym is a standardized framework for evaluating and training data science agents, addressing shortcomings of existing benchmarks. It offers a holistic, data-grounded task suite and enables execution-verified agent training. This allows rigorous measurement of agents' analytical capabilities, dem...

🔹 Publication Date: Published on Jan 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.16344
• PDF: https://arxiv.org/pdf/2601.16344

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#DataScience #AI #MachineLearning #AIagents #Research
This media is not supported in your browser
VIEW IN TELEGRAM
Memory-V2V: Augmenting Video-to-Video Diffusion Models with Memory

📝 Summary:
Memory-V2V enhances multi-turn video editing by adding explicit memory to diffusion models. It ensures cross-consistency using efficient token compression and retrieval. This significantly improves video consistency and performance with low computational cost.

🔹 Publication Date: Published on Jan 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.16296
• PDF: https://arxiv.org/pdf/2601.16296
• Project Page: https://dohunlee1.github.io/MemoryV2V

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#VideoEditing #DiffusionModels #GenerativeAI #ComputerVision #MachineLearning
SWE-Pruner: Self-Adaptive Context Pruning for Coding Agents

📝 Summary:
SWE-Pruner is a self-adaptive context pruning framework for coding agents. It performs task-aware adaptive pruning, guided by explicit agent goals and a neural skimmer, to reduce long context token usage by 23-54 percent with minimal performance loss.

🔹 Publication Date: Published on Jan 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.16746
• PDF: https://arxiv.org/pdf/2601.16746
• Github: https://github.com/Ayanami1314/swe-pruner

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#AIAgents #ContextPruning #LLM #AI #SoftwareEngineering
Inference-Time Scaling of Verification: Self-Evolving Deep Research Agents via Test-Time Rubric-Guided Verification

📝 Summary:
A self-evolving framework improves Deep Research Agents via inference-time, rubric-guided verification. This method iteratively refines outputs without retraining, achieving 8-11% accuracy gains with the DeepVerifier system and releasing a verification dataset.

🔹 Publication Date: Published on Jan 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.15808
• PDF: https://arxiv.org/pdf/2601.15808

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#AI #MachineLearning #DeepLearning #Verification #SelfEvolvingAI
MeepleLM: A Virtual Playtester Simulating Diverse Subjective Experiences

📝 Summary:
MeepleLM is an AI virtual playtester providing constructive critique for board game design by simulating diverse player experiences. It models subjective feedback via persona-specific reasoning, outperforming commercial AI in critique quality and community alignment.

🔹 Publication Date: Published on Jan 12

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.07251
• PDF: https://arxiv.org/pdf/2601.07251
• Github: https://github.com/leroy9472/MeepleLM

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#AI #GameDesign #BoardGames #Simulation #LLM
SALAD: Achieve High-Sparsity Attention via Efficient Linear Attention Tuning for Video Diffusion Transformer

📝 Summary:
SALAD improves video Diffusion Transformers by combining linear and sparse attention with an input-dependent gating mechanism. It achieves 90% sparsity and a 1.72x speedup while maintaining quality and requiring minimal finetuning data.

🔹 Publication Date: Published on Jan 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.16515
• PDF: https://arxiv.org/pdf/2601.16515

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#VideoDiffusion #Transformers #Sparsity #EfficientAI #DeepLearning
Mecellem Models: Turkish Models Trained from Scratch and Continually Pre-trained for the Legal Domain

📝 Summary:
Mecellem models are a framework for specialized Turkish legal language models. They feature a scratch-trained encoder achieving top retrieval rankings with efficiency, and a continually pre-trained decoder for legal domain adaptation, reducing legal text perplexity.

🔹 Publication Date: Published on Jan 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.16018
• PDF: https://arxiv.org/pdf/2601.16018
• Project Page: https://huggingface.co/collections/newmindai/mecellem-models
• Github: https://github.com/newmindai/mecellem-models

🔹 Models citing this paper:
https://huggingface.co/newmindai/Mursit-Base-TR-Retrieval
https://huggingface.co/newmindai/Mursit-Base
https://huggingface.co/newmindai/Mursit-Large-TR-Retrieval

Datasets citing this paper:
https://huggingface.co/datasets/newmindai/caselaw-retrieval
https://huggingface.co/datasets/newmindai/contract-retrieval
https://huggingface.co/datasets/newmindai/regulation-retrieval

Spaces citing this paper:
https://huggingface.co/spaces/newmindai/Mizan

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#LegalAI #TurkishNLP #LLM #InformationRetrieval #DomainAdaptation
Jet-RL: Enabling On-Policy FP8 Reinforcement Learning with Unified Training and Rollout Precision Flow

📝 Summary:
Quantized RL faces instability using FP8 rollout with BF16 training. Jet-RL proposes a unified FP8 precision for both training and rollout. This minimizes numerical mismatch, achieving stable convergence and significant speedups.

🔹 Publication Date: Published on Jan 20

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.14243
• PDF: https://arxiv.org/pdf/2601.14243

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research
Guidelines to Prompt Large Language Models for Code Generation: An Empirical Characterization

📝 Summary:
Research derives and evaluates prompt optimization guidelines for code generation tasks in software engineering, identifying 10 specific improvement patterns related to input/output specification, con...

🔹 Publication Date: Published on Jan 19

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.13118
• PDF: https://arxiv.org/pdf/2601.13118

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research
1👍1
Knowledge is Not Enough: Injecting RL Skills for Continual Adaptation

📝 Summary:
LLMs struggle to apply new knowledge effectively via SFT alone. PaST combines SFT with injecting a domain-agnostic Skill Vector, derived from RL, to efficiently transfer reasoning skills. This novel framework significantly improves performance in question answering and tool-use tasks.

🔹 Publication Date: Published on Jan 16

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.11258
• PDF: https://arxiv.org/pdf/2601.11258

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#LLM #ReinforcementLearning #ContinualLearning #AI #MachineLearning
1
Dancing in Chains: Strategic Persuasion in Academic Rebuttal via Theory of Mind

📝 Summary:
RebuttalAgent is a novel AI framework that applies Theory of Mind to academic rebuttal. It models reviewer mental states to formulate strategic, persuasive responses, significantly outperforming existing models.

🔹 Publication Date: Published on Jan 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.15715
• PDF: https://arxiv.org/pdf/2601.15715

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#AI #TheoryOfMind #AcademicRebuttal #NLP #MachineLearning
2
GameTalk: Training LLMs for Strategic Conversation

📝 Summary:
The GameTalk framework trains large language models for strategic multi-turn dialogue, optimizing global objectives using whole-conversation reward signals. This approach significantly outperforms untrained models, showing conversational fine-tuning is a promising path for LLM reasoning and negot...

🔹 Publication Date: Published on Jan 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.16276
• PDF: https://arxiv.org/pdf/2601.16276

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#LLMs #ConversationalAI #StrategicDialogue #AITraining #AIReasoning
1
ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch

📝 Summary:
ChartVerse is a framework that synthesizes complex charts and reliable reasoning data for VLMs. It uses a novel metric, Rollout Posterior Entropy, for complexity-aware chart generation and an answer-first QA synthesis to ensure reasoning rigor. This leads to state-of-the-art performance in chart ...

🔹 Publication Date: Published on Jan 20

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.13606
• PDF: https://arxiv.org/pdf/2601.13606
• Project Page: https://chartverse.github.io/
• Github: https://github.com/starriver030515/ChartVerse

🔹 Models citing this paper:
https://huggingface.co/opendatalab/ChartVerse-Coder
https://huggingface.co/opendatalab/ChartVerse-2B
https://huggingface.co/opendatalab/ChartVerse-8B

Datasets citing this paper:
https://huggingface.co/datasets/opendatalab/ChartVerse-SFT-1800K
https://huggingface.co/datasets/opendatalab/ChartVerse-SFT-600K
https://huggingface.co/datasets/opendatalab/ChartVerse-RL-40K

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

For more data science resources:
https://news.1rj.ru/str/DataScienceT

#AI #VLMs #ChartReasoning #MachineLearning #DataScience
1
Do you see yourself as a programmer, researcher, or engineer?
Anonymous Poll
44%
Programmer
23%
Researcher
33%
Engineer