✨Transformers converge to invariant algorithmic cores

📝 Summary:
Independently trained transformers converge to shared low-dimensional algorithmic cores. These compact invariants reveal the computational essence across training runs and scales, suggesting a new focus for mechanistic interpretability.

🔹 Publication Date: Published on Feb 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.22600
• PDF: https://arxiv.org/pdf/2602.22600

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Humans and LLMs Diverge on Probabilistic Inferences

📝 Summary:
LLMs consistently fail to replicate human probabilistic reasoning patterns in open-ended inferences, despite strong performance on logical and mathematical tasks. The ProbCOPA dataset reveals this key divergence, highlighting a need to evaluate AI reasoning beyond deterministic settings.

🔹 Publication Date: Published on Feb 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.23546
• PDF: https://arxiv.org/pdf/2602.23546
• Project Page: https://grvkamath.github.io/probcopa-demo/index.html
• Github: https://github.com/McGill-NLP/probabilistic-reasoning

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Easy to Learn, Yet Hard to Forget: Towards Robust Unlearning Under Bias

📝 Summary:
Machine unlearning in biased models suffers from shortcut unlearning, where bias attributes are forgotten instead of class attributes. The CUPID framework addresses this by partitioning data based on loss sharpness and disentangling causal and bias pathways for targeted updates, achieving state-o...

🔹 Publication Date: Published on Feb 25

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.21773
• PDF: https://arxiv.org/pdf/2602.21773

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨QEDBENCH: Quantifying the Alignment Gap in Automated Evaluation of University-Level Mathematical Proofs

📝 Summary:
QEDBench quantifies a systematic Alignment Gap in LLM evaluation of university-level math proofs. It reveals score inflation in some frontier models and a critical reasoning gap in discrete mathematics for others. QEDBench is now a public benchmark.

🔹 Publication Date: Published on Feb 24

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.20629
• PDF: https://arxiv.org/pdf/2602.20629
• Project Page: https://quanquancliu.com/QEDBench/index.html
• Github: https://github.com/qqliu/Yale-QEDBench

✨ Datasets citing this paper:
• https://huggingface.co/datasets/qqggez/QEDBench

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨ParEVO: Synthesizing Code for Irregular Data: High-Performance Parallelism through Agentic Evolution

📝 Summary:
ParEVO synthesizes high-performance parallel algorithms for irregular data structures using specialized LLMs and evolutionary coding to achieve significant speedups over existing methods. AI-generated...

🔹 Publication Date: Published on Mar 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.02510
• PDF: https://arxiv.org/pdf/2603.02510
• Project Page: https://quanquancliu.com/ParEVO/index.html
• Github: https://github.com/WildAlg/ParEVO

🔹 Models citing this paper:
• https://huggingface.co/qqggez/qwen3-30b-sft-stage2-merged
• https://huggingface.co/qqggez/deepseek-parlay-6.7b

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨LFPO: Likelihood-Free Policy Optimization for Masked Diffusion Models

📝 Summary:
LFPO trains diffusion LLMs by directly optimizing denoising logits through geometric velocity rectification, avoiding complex likelihood calculations. This yields precise gradient estimates, faster inference, and better performance on code and reasoning tasks.

🔹 Publication Date: Published on Mar 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.01563
• PDF: https://arxiv.org/pdf/2603.01563
• Github: https://github.com/kithib/LFPO

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Words & Weights: Streamlining Multi-Turn Interactions via Co-Adaptation

📝 Summary:
ROSA2 framework addresses multi-turn LLM interactions by jointly optimizing textual intent clarity and model parameters, achieving superior mathematical reasoning performance with fewer interaction ro...

🔹 Publication Date: Published on Mar 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.01375
• PDF: https://arxiv.org/pdf/2603.01375
• Github: https://github.com/kithib/ROSA2

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Helios: Real Real-Time Long Video Generation Model

📝 Summary:
Helios is a 14B autoregressive diffusion model that achieves real-time minute-scale video generation at 19.5 FPS on a single GPU. It innovatively overcomes long-video drifting and real-time performance challenges without conventional acceleration or anti-drifting techniques. Helios supports T2V, ...

🔹 Publication Date: Published on Mar 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04379
• PDF: https://arxiv.org/pdf/2603.04379
• Github: https://pku-yuangroup.github.io/Helios-Page/

🔹 Models citing this paper:
• https://huggingface.co/BestWishYsh/Helios-Base
• https://huggingface.co/BestWishYsh/Helios-Distilled
• https://huggingface.co/BestWishYsh/Helios-Mid

✨ Spaces citing this paper:
• https://huggingface.co/spaces/multimodalart/Helios-Distilled

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨ArtHOI: Articulated Human-Object Interaction Synthesis by 4D Reconstruction from Video Priors

📝 Summary:
ArtHOI synthesizes articulated human-object interactions by formulating 4D reconstruction from monocular video priors, using optical flow for part segmentation and a decoupled reconstruction pipeline ...

🔹 Publication Date: Published on Mar 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04338
• PDF: https://arxiv.org/pdf/2603.04338
• Project Page: https://arthoi.github.io/
• Github: https://github.com/Inso-13/ArtHOI

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Experience Memory

📝 Summary:
A memory mechanism called Memex enables large language model agents to handle long-horizon tasks more effectively by maintaining compact context through structured summaries while storing full interac...

🔹 Publication Date: Published on Mar 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04257
• PDF: https://arxiv.org/pdf/2603.04257

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research