ML Research Hub
32.9K subscribers
5.35K photos
332 videos
24 files
5.78K links
Advancing research in Machine Learning – practical insights, tools, and techniques for researchers.

Admin: @HusseinSheikho || @Hussein_Sheikho
Download Telegram
HalluHard: A Hard Multi-Turn Hallucination Benchmark

📝 Summary:
Large language models continue to generate plausible but ungrounded factual claims in multi-turn dialogue, with hallucinations remaining significant even when utilizing web search for verification acr...

🔹 Publication Date: Published on Feb 1

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.01031
• PDF: https://arxiv.org/pdf/2602.01031

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Trust The Typical

📝 Summary:
Trust The Typical T3 frames LLM safety as an out-of-distribution detection problem, learning what is safe in semantic space. It achieves state-of-the-art performance without harmful example training, drastically reducing false positives and generalizing across languages with low overhead.

🔹 Publication Date: Published on Feb 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.04581
• PDF: https://arxiv.org/pdf/2602.04581

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Learning to Repair Lean Proofs from Compiler Feedback

📝 Summary:
A new dataset, APRIL, pairs erroneous Lean proofs with compiler feedback, corrected proofs, and natural language diagnoses. Training language models on APRIL substantially improves proof repair accuracy and feedback-conditioned reasoning, outperforming existing baselines.

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.02990
• PDF: https://arxiv.org/pdf/2602.02990

Datasets citing this paper:
https://huggingface.co/datasets/uw-math-ai/APRIL

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
MeKi: Memory-based Expert Knowledge Injection for Efficient LLM Scaling

📝 Summary:
MeKi enables efficient large language model deployment on edge devices by injecting pre-stored semantic knowledge through token-level memory experts and re-parameterization techniques. AI-generated su...

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.03359
• PDF: https://arxiv.org/pdf/2602.03359

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Semantic Search over 9 Million Mathematical Theorems

📝 Summary:
Large-scale semantic theorem retrieval system demonstrates superior performance over existing baselines using a 9.2 million theorem corpus with systematic analysis of representation context, language ...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05216
• PDF: https://arxiv.org/pdf/2602.05216

Datasets citing this paper:
https://huggingface.co/datasets/uw-math-ai/theorem-search-dataset

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
RISE-Video: Can Video Generators Decode Implicit World Rules?

📝 Summary:
RISE-Video presents a novel benchmark for evaluating text-image-to-video synthesis models based on cognitive reasoning rather than visual fidelity, using a multi-dimensional metric system and automate...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05986
• PDF: https://arxiv.org/pdf/2602.05986
• Github: https://github.com/VisionXLab/Rise-Video

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Length-Unbiased Sequence Policy Optimization: Revealing and Controlling Response Length Variation in RLVR

📝 Summary:
Research analyzes RLVR algorithms' impact on response length in LLMs and VLMs, proposing LUSPO to eliminate length bias and improve reasoning performance. AI-generated summary Recent applications of R...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05261
• PDF: https://arxiv.org/pdf/2602.05261
• Github: https://github.com/murphy4122/LUSPO

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
SwimBird: Eliciting Switchable Reasoning Mode in Hybrid Autoregressive MLLMs

📝 Summary:
SwimBird is a reasoning-switchable multimodal large language model that dynamically selects between text-only, vision-only, and interleaved vision-text reasoning modes based on input queries, achievin...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.06040
• PDF: https://arxiv.org/pdf/2602.06040
• Project Page: https://accio-lab.github.io/SwimBird
• Github: https://github.com/Accio-Lab/SwimBird

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Grounding and Enhancing Informativeness and Utility in Dataset Distillation

📝 Summary:
Dataset distillation method that balances informativeness and utility through game-theoretic and gradient-based optimization techniques, achieving improved performance on ImageNet-1K. AI-generated sum...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21296
• PDF: https://arxiv.org/pdf/2601.21296

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Reinforcement World Model Learning for LLM-based Agents

📝 Summary:
Reinforcement World Model Learning enables LLM-based agents to better anticipate action consequences and adapt to environment dynamics through self-supervised training that aligns simulated and real-w...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05842
• PDF: https://arxiv.org/pdf/2602.05842

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Breaking the Static Graph: Context-Aware Traversal for Robust Retrieval-Augmented Generation

📝 Summary:
CatRAG addresses limitations in retrieval-augmented generation by introducing a query-adaptive framework that improves multi-hop reasoning through symbolic anchoring, dynamic edge weighting, and key-f...

🔹 Publication Date: Published on Feb 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.01965
• PDF: https://arxiv.org/pdf/2602.01965

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Context Forcing: Consistent Autoregressive Video Generation with Long Context

📝 Summary:
Context Forcing addresses student-teacher mismatch in long video generation by using a long-context teacher to guide long-rollout students through a Slow-Fast Memory architecture that extends context ...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.06028
• PDF: https://arxiv.org/pdf/2602.06028
• Project Page: https://chenshuo20.github.io/Context_Forcing/
• Github: https://github.com/TIGER-AI-Lab/Context-Forcing

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Late-to-Early Training: LET LLMs Learn Earlier, So Faster and Better

📝 Summary:
Large language models can be trained more efficiently by transferring knowledge from later training phases to earlier layers during initial training, achieving faster convergence and improved performa...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05393
• PDF: https://arxiv.org/pdf/2602.05393

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
LatentMem: Customizing Latent Memory for Multi-Agent Systems

📝 Summary:
LatentMem is a learnable multi-agent memory framework that customizes agent-specific memories through latent representations, improving performance in multi-agent systems without modifying underlying ...

🔹 Publication Date: Published on Feb 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.03036
• PDF: https://arxiv.org/pdf/2602.03036
• Github: https://github.com/KANABOON1/LatentMem

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
ProAct: Agentic Lookahead in Interactive Environments

📝 Summary:
ProAct enhances LLM agents' long-horizon planning by combining supervised fine-tuning with search-derived trajectories and a Monte-Carlo critic for improved policy optimization. AI-generated summary E...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05327
• PDF: https://arxiv.org/pdf/2602.05327
• Github: https://github.com/GreatX3/ProAct

🔹 Models citing this paper:
https://huggingface.co/biang889/ProAct

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Media is too big
VIEW IN TELEGRAM
FastVMT: Eliminating Redundancy in Video Motion Transfer

📝 Summary:
FastVMT accelerates video motion transfer by addressing computational redundancies in Diffusion Transformer architecture through localized attention masking and gradient reuse optimization. AI-generat...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05551
• PDF: https://arxiv.org/pdf/2602.05551
• Project Page: https://fastvmt.github.io/

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Retrieval-Infused Reasoning Sandbox: A Benchmark for Decoupling Retrieval and Reasoning Capabilities

📝 Summary:
DeR2 presents a controlled evaluation framework for assessing language models' document-grounded reasoning capabilities by isolating reasoning from retrieval and toolchain decisions. AI-generated summ...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21937
• PDF: https://arxiv.org/pdf/2601.21937
• Project Page: https://huggingface.co/m-a-p
• Github: https://retrieval-infused-reasoning-sandbox.github.io/

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Pathwise Test-Time Correction for Autoregressive Long Video Generation

📝 Summary:
Test-Time Correction addresses error accumulation in distilled autoregressive diffusion models for long-video synthesis by using initial frames as reference anchors to calibrate stochastic states duri...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05871
• PDF: https://arxiv.org/pdf/2602.05871

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
BABE: Biology Arena BEnchmark

📝 Summary:
BABE is a biology-focused benchmark designed to evaluate AI systems' ability to perform experimental reasoning and causal inference similar to practicing scientists. AI-generated summary The rapid evo...

🔹 Publication Date: Published on Feb 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.05857
• PDF: https://arxiv.org/pdf/2602.05857

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
UniAudio 2.0: A Unified Audio Language Model with Text-Aligned Factorized Audio Tokenization

📝 Summary:
Researchers developed a discrete audio codec called ReasoningCodec that separates audio into reasoning and reconstruction tokens for improved understanding and generation, and created UniAudio 2.0, a ...

🔹 Publication Date: Published on Feb 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.04683
• PDF: https://arxiv.org/pdf/2602.04683
• Project Page: https://dongchaoyang.top/UniAudio2Demo/
• Github: https://github.com/yangdongchao/UniAudio2

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
1
Steering LLMs via Scalable Interactive Oversight

📝 Summary:
Scalable Interactive Oversight framework decomposes complex tasks into manageable decision trees to enhance human supervision and alignment in AI systems. AI-generated summary As Large Language Models...

🔹 Publication Date: Published on Feb 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.04210
• PDF: https://arxiv.org/pdf/2602.04210

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research