ML Research Hub
32.9K subscribers
4.72K photos
292 videos
24 files
5.1K links
Advancing research in Machine Learning – practical insights, tools, and techniques for researchers.

Admin: @HusseinSheikho || @Hussein_Sheikho
Download Telegram
Training Reasoning Models on Saturated Problems via Failure-Prefix Conditioning

📝 Summary:
Reinforcement learning training stalls on saturated problems as informative failures are hard to find. Failure-prefix conditioning addresses this by training on prefixes from rare incorrect reasoning paths, exposing models to failures. This boosts performance, maintains efficiency, and improves r...

🔹 Publication Date: Published on Jan 28

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.20829
• PDF: https://arxiv.org/pdf/2601.20829
• Github: https://github.com/minwukim/training-on-saturated-problems

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

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

#ReinforcementLearning #MachineLearning #ArtificialIntelligence #DeepLearning #AIResearch
1
MM-Agent: LLM as Agents for Real-world Mathematical Modeling Problem

📝 Summary:
MM-Agent is an expert-inspired framework that enables LLMs to excel in real-world mathematical modeling by decomposing the task into four stages. It significantly outperforms human experts and baseline agents on a new benchmark, proving its practical effectiveness as a modeling copilot.

🔹 Publication Date: Published on May 20, 2025

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2505.14148
• PDF: https://arxiv.org/pdf/2505.14148
• Github: https://github.com/usail-hkust/llm-mm-agent

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

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

#LLM #MathematicalModeling #AIAgents #ArtificialIntelligence #DataScience
1
VERGE: Formal Refinement and Guidance Engine for Verifiable LLM Reasoning

📝 Summary:
VERGE is a neurosymbolic framework that combines LLMs with SMT solvers for verification-guided iterative refinement of reasoning. It enhances logical correctness through formal semantic checking, semantic routing, and precise error localization, achieving an 18.7% performance uplift on reasoning ...

🔹 Publication Date: Published on Jan 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.20055
• PDF: https://arxiv.org/pdf/2601.20055

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

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

#LLM #NeurosymbolicAI #FormalVerification #AIReasoning #SMTSolvers
2🔥1
Group Distributionally Robust Optimization-Driven Reinforcement Learning for LLM Reasoning

📝 Summary:
This paper introduces Multi-Adversary GDRO to improve LLM reasoning. It dynamically adapts training distributions by classifying prompt difficulty and reallocating resources. This boosts accuracy by over 10% compared to GRPO, focusing compute on hard problems.

🔹 Publication Date: Published on Jan 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.19280
• PDF: https://arxiv.org/pdf/2601.19280

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

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

#LLMReasoning #ReinforcementLearning #Optimization #MachineLearning #AI
1
Persona Prompting as a Lens on LLM Social Reasoning

📝 Summary:
Persona prompting improves LLM classification on subjective tasks like hate speech but degrades explanation quality. It fails to mitigate demographic biases and align with real-world personas, as models remain resistant to significant steering and over-flag content as harmful. This reveals a crit...

🔹 Publication Date: Published on Jan 28

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.20757
• PDF: https://arxiv.org/pdf/2601.20757
• Github: https://github.com/jingyng/PP-social-reasoning

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

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

#LLM #PersonaPrompting #BiasInAI #AIethics #NLP
🔥1
How AI Impacts Skill Formation

📝 Summary:
AI assistance impairs skill acquisition for novice workers, hindering conceptual understanding and debugging. Heavy AI reliance is not a shortcut to competence. Careful AI adoption is crucial to preserve skill formation.

🔹 Publication Date: Published on Jan 28

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.20245
• PDF: https://arxiv.org/pdf/2601.20245
• Project Page: https://www.anthropic.com/research/AI-assistance-coding-skills

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

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

#AI #SkillFormation #WorkforceDevelopment #LearningScience #HumanAICollaboration
FP8-RL: A Practical and Stable Low-Precision Stack for LLM Reinforcement Learning

📝 Summary:
FP8-RL presents a practical FP8 rollout stack for LLM reinforcement learning, addressing computational and memory bottlenecks. It employs blockwise FP8, KV-cache recalibration, and importance sampling to mitigate train-inference mismatch. This achieves up to 44% throughput gains while preserving ...

🔹 Publication Date: Published on Jan 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.18150
• PDF: https://arxiv.org/pdf/2601.18150

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

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

#LLM #ReinforcementLearning #FP8 #MachineLearning #AIResearch
Language-based Trial and Error Falls Behind in the Era of Experience

📝 Summary:
LLMs struggle in nonlinguistic tasks due to costly exploration. SCOUT uses lightweight scouts for efficient exploration, then fine-tunes LLMs via SFT and RL. This boosts performance and saves GPU hours, outperforming proprietary models.

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21754
• PDF: https://arxiv.org/pdf/2601.21754
• Project Page: https://scout-cs.github.io/
• Github: https://github.com/Harry-mic/SCOUT

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Llama-3.1-FoundationAI-SecurityLLM-Reasoning-8B Technical Report

📝 Summary:
A two-stage trained cybersecurity reasoning model achieves competitive performance on specialized tasks while maintaining general capabilities through supervised fine-tuning and reinforcement learning...

🔹 Publication Date: Published on Jan 28

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21051
• PDF: https://arxiv.org/pdf/2601.21051
• Project Page: https://huggingface.co/fdtn-ai/Foundation-Sec-8B-Reasoning

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
VTC-R1: Vision-Text Compression for Efficient Long-Context Reasoning

📝 Summary:
VTC-R1 enables efficient long-context reasoning by compressing textual traces into compact images and iteratively feeding them back into vision-language models as optical memory, achieving significant...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22069
• PDF: https://arxiv.org/pdf/2601.22069
• Github: https://github.com/w-yibo/VTC-R1

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Exploring Reasoning Reward Model for Agents

📝 Summary:
Agent-RRM, a multi-faceted reward model, provides structured feedback for agentic trajectories through reasoning traces, critiques, and performance scores, with unified feedback integration showing su...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/pdf/2601.22154
• PDF: https://arxiv.org/pdf/2601.22154
• Github: https://github.com/kxfan2002/Reagent

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Beyond Imitation: Reinforcement Learning for Active Latent Planning

📝 Summary:
Active latent planning method improves reasoning accuracy and efficiency by modeling latent token supervision as conditional VAE and using reinforcement learning with coherence rewards. AI-generated s...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21598
• PDF: https://arxiv.org/pdf/2601.21598

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Generation Enhances Understanding in Unified Multimodal Models via Multi-Representation Generation

📝 Summary:
UniMRG enhances unified multimodal models by training them to generate multiple visual representations, improving both understanding and generation capabilities through complementary information captu...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21406
• PDF: https://arxiv.org/pdf/2601.21406
• Github: https://github.com/Sugewud/UniMRG

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
WebArbiter: A Principle-Guided Reasoning Process Reward Model for Web Agents

📝 Summary:
WebArbiter introduces a reasoning-first WebPRM that formulates reward modeling as text generation to improve web navigation through structured justifications and preference verdicts, outperforming exi...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21872
• PDF: https://arxiv.org/pdf/2601.21872

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Media is too big
VIEW IN TELEGRAM
DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation

📝 Summary:
DynamicVLA addresses dynamic object manipulation challenges through a compact vision-language-action model with temporal reasoning and closed-loop adaptation, supported by a new benchmark for dynamic ...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22153
• PDF: https://arxiv.org/pdf/2601.22153
• Project Page: https://haozhexie.com/project/dynamic-vla
• Github: https://github.com/hzxie/DynamicVLA

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods

📝 Summary:
A large-scale multimodal reasoning dataset called MMFineReason is introduced to improve vision language models' performance through high-quality reasoning annotations and demonstrates superior paramet...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21821
• PDF: https://arxiv.org/pdf/2601.21821
• Project Page: https://mmfinereason.github.io/

🔹 Models citing this paper:
https://huggingface.co/OpenDataArena/MMFineReason-8B
https://huggingface.co/OpenDataArena/MMFineReason-4B
https://huggingface.co/OpenDataArena/MMFineReason-2B

Datasets citing this paper:
https://huggingface.co/datasets/OpenDataArena/MMFineReason-1.8M
https://huggingface.co/datasets/OpenDataArena/MMFineReason-SFT-123K

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
MAD: Modality-Adaptive Decoding for Mitigating Cross-Modal Hallucinations in Multimodal Large Language Models

📝 Summary:
Multimodal Large Language Models suffer from cross-modal hallucinations where one modality incorrectly influences generation from another, leading to fabricated outputs; this exposes a fundamental def...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21181
• PDF: https://arxiv.org/pdf/2601.21181

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
Qwen3-ASR Technical Report

📝 Summary:
The Qwen3-ASR family introduces speech recognition models with language identification capabilities and a non-autoregressive forced alignment model, achieving state-of-the-art performance and efficien...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21337
• PDF: https://arxiv.org/pdf/2601.21337

🔹 Models citing this paper:
https://huggingface.co/Qwen/Qwen3-ASR-1.7B
https://huggingface.co/Qwen/Qwen3-ASR-0.6B
https://huggingface.co/Qwen/Qwen3-ForcedAligner-0.6B

Spaces citing this paper:
https://huggingface.co/spaces/Qwen/Qwen3-ASR
https://huggingface.co/spaces/prithivMLmods/Qwen3-TTS-Daggr-UI
https://huggingface.co/spaces/sxjeru/Qwen3-ASR-1.7B

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
ConceptMoE: Adaptive Token-to-Concept Compression for Implicit Compute Allocation

📝 Summary:
ConceptMoE dynamically allocates computation by merging similar tokens into concept representations, improving both performance and efficiency in large language models through adaptive processing and ...

🔹 Publication Date: Published on Jan 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.21420
• PDF: https://arxiv.org/pdf/2601.21420
• Github: https://github.com/ZihaoHuang-notabot/ConceptMoE

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research
1