🚀 Sber has released two open-source MoE models: GigaChat-3.1 Ultra and Lightning

Both code and weights are available under the MIT license on HuggingFace.

👉 Key details:

• Trained from scratch (not a finetune) on proprietary data and infrastructure
• Mixture-of-Experts (MoE) architecture

Models:

🧠 GigaChat-3.1 Ultra
• 702B MoE model for high-performance environments
• Outperforms DeepSeek-V3-0324 and Qwen3-235B on math and reasoning benchmarks
• Supports FP8 training and MTP

⚡️ GigaChat-3.1 Lightning
• 10B model (1.8B active parameters)
• Outperforms Qwen3-4B and Gemma-3-4B on Sber benchmarks
• Efficient local inference
• Up to 256k context

Engineering highlights:

• Custom metric to detect and reduce generation loops
• DPO training moved to native FP8
• Improvements in post-training pipeline
• Identified and fixed a critical issue affecting evaluation quality

🌍 Trained on 14 languages (optimized for English and Russian)

Use cases:

• chatbots
• AI assistants
• copilots
• internal ML systems

Sber provides a solid open foundation for developers to build production-ready AI systems with lower infrastructure costs.

✨QiMeng-PRepair: Precise Code Repair via Edit-Aware Reward Optimization

📝 Summary:
PRepair tackles over-editing in AI program repair by maximizing correct code reuse. It combines controlled bug injection and edit-aware policy optimization using an edit-aware reward. This framework significantly improves repair precision and decoding throughput.

🔹 Publication Date: Published on Apr 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.05963
• PDF: https://arxiv.org/pdf/2604.05963
• Github: https://github.com/kcxain/QiMeng-PRepair

🔹 Models citing this paper:
• https://huggingface.co/kcxain/Prepair-Python-7B-EA
• https://huggingface.co/kcxain/Prepair-Verilog-7B-EA

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

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

#ProgramRepair #AI #MachineLearning #ReinforcementLearning #SoftwareEngineering

✨Watch Before You Answer: Learning from Visually Grounded Post-Training

📝 Summary:
VLMs struggle with video understanding due to text biases in benchmarks and training data. VidGround uses only visually grounded questions for post-training to eliminate these biases. This improves VLM performance and emphasizes the need for high-quality, visually grounded data.

🔹 Publication Date: Published on Apr 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.05117
• PDF: https://arxiv.org/pdf/2604.05117
• Project Page: https://vidground.etuagi.com
• Github: https://github.com/reacher-z/vidground

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

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

#VLMs #VideoUnderstanding #AI #MachineLearning #ComputerVision

✨DARE: Diffusion Large Language Models Alignment and Reinforcement Executor

📝 Summary:
Diffusion large language models are gaining attention as alternatives to autoregressive models, utilizing iterative denoising and parallel generation instead of sequential token processing, yet their ...

🔹 Publication Date: Published on Apr 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.04215
• PDF: https://arxiv.org/pdf/2604.04215
• Github: https://github.com/yjyddq/DARE

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Context-Value-Action Architecture for Value-Driven Large Language Model Agents

📝 Summary:
LLMs show rigid, polarized behavior worsening with reasoning. The Context-Value-Action CVA architecture decouples actions from reasoning using a human-data Value Verifier, mitigating polarization and improving behavioral fidelity.

🔹 Publication Date: Published on Apr 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.05939
• PDF: https://arxiv.org/pdf/2604.05939

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Can Natural Image Autoencoders Compactly Tokenize fMRI Volumes for Long-Range Dynamics Modeling?

📝 Summary:
TABLeT uses a 2D natural image autoencoder to tokenize fMRI volumes into compact continuous tokens, enabling efficient long-sequence spatiotemporal modeling with a simple Transformer encoder while mai...

🔹 Publication Date: Published on Apr 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.03619
• PDF: https://arxiv.org/pdf/2604.03619
• Project Page: https://concarne2.github.io/tablet_project_page/
• Github: https://github.com/beotborry/TABLeT

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Squeez: Task-Conditioned Tool-Output Pruning for Coding Agents

📝 Summary:
A task-conditioned tool-output pruning model effectively reduces input tokens for coding agents. It achieves 0.86 recall and 0.80 F1, removing 92% of tokens, outperforming larger zero-shot models and heuristic baselines.

🔹 Publication Date: Published on Apr 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.04979
• PDF: https://arxiv.org/pdf/2604.04979
• Github: https://github.com/KRLabsOrg/squeez

🔹 Models citing this paper:
• https://huggingface.co/KRLabsOrg/squeez-2b

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

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

#CodingAgents #LLM #TokenPruning #AI #MachineLearning

✨General Multimodal Protein Design Enables DNA-Encoding of Chemistry

📝 Summary:
DISCO is a multimodal deep generative model that co-designs protein sequences and 3D structures to create novel heme enzymes with unprecedented catalytic capabilities. AI-generated summary Evolution i...

🔹 Publication Date: Published on Apr 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.05181
• PDF: https://arxiv.org/pdf/2604.05181
• Project Page: https://disco-design.github.io/
• Github: https://github.com/DISCO-design/DISCO

🔹 Models citing this paper:
• https://huggingface.co/DISCO-Design/DISCO

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Expert-Choice Routing Enables Adaptive Computation in Diffusion Language Models

📝 Summary:
Expert-choice routing improves diffusion language model mixture-of-experts by providing deterministic load balancing and adaptive computation allocation based on denoising steps. AI-generated summary ...

🔹 Publication Date: Published on Apr 2

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.01622
• PDF: https://arxiv.org/pdf/2604.01622
• Github: https://github.com/zhangshuibai/EC-DLM

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨ClawsBench: Evaluating Capability and Safety of LLM Productivity Agents in Simulated Workspaces

📝 Summary:
ClawsBench evaluates LLM productivity agents in realistic workflows with mock services, assessing capability and safety. It shows agents achieve 39-64% task success but also 7-33% unsafe actions, identifying recurring patterns.

🔹 Publication Date: Published on Apr 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.05172
• PDF: https://arxiv.org/pdf/2604.05172
• Project Page: https://clawsbench.com/
• Github: https://github.com/benchflow-ai/ClawsBench

✨ Datasets citing this paper:
• https://huggingface.co/datasets/benchflow/ClawsBench

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

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

#LLM #AIAgents #AISafety #Benchmarking #AIResearch

✨CUE-R: Beyond the Final Answer in Retrieval-Augmented Generation

📝 Summary:
Researchers developed a framework to measure the operational utility of individual retrieved items in retrieval-augmented generation systems by perturbing evidence and analyzing changes in correctness...

🔹 Publication Date: Published on Apr 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.05467
• PDF: https://arxiv.org/pdf/2604.05467
• Github: https://github.com/jainsid24/cue-r

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨REAM: Merging Improves Pruning of Experts in LLMs

📝 Summary:
Router-weighted Expert Activation Merging (REAM) is proposed as a novel method for reducing memory requirements in Mixture-of-Experts large language models by grouping and merging expert weights inste...

🔹 Publication Date: Published on Apr 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.04356
• PDF: https://arxiv.org/pdf/2604.04356
• Project Page: https://bknyaz.github.io/blog/2026/moe/
• Github: https://github.com/SamsungSAILMontreal/ream

🔹 Models citing this paper:
• https://huggingface.co/bknyaz/Qwen3-Coder-Next-REAM
• https://huggingface.co/SamsungSAILMontreal/Qwen3-30B-A3B-Instruct-2507-REAM
• https://huggingface.co/bknyaz/Qwen3-Next-80B-A3B-Instruct-REAM

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Personalized RewardBench: Evaluating Reward Models with Human Aligned Personalization

📝 Summary:
Personalized RewardBench evaluates reward models' ability to capture individual user preferences, revealing significant challenges in current models and demonstrating superior correlation with downstr...

🔹 Publication Date: Published on Apr 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.07343
• PDF: https://arxiv.org/pdf/2604.07343
• Project Page: https://huggingface.co/datasets/QiyaoMa/Personalized-RewardBench
• Github: https://github.com/Martin-qyma/Personalized-RewardBench

✨ Datasets citing this paper:
• https://huggingface.co/datasets/QiyaoMa/Personalized-RewardBench

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨MARS: Enabling Autoregressive Models Multi-Token Generation

📝 Summary:
MARS is a fine-tuning method that enables autoregressive language models to predict multiple tokens per forward pass without architectural changes, maintaining accuracy while improving throughput and ...

🔹 Publication Date: Published on Apr 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.07023
• PDF: https://arxiv.org/pdf/2604.07023

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨MoRight: Motion Control Done Right

📝 Summary:
MoRight is a unified framework that enables disentangled motion control and causal relationship modeling in video generation, allowing separate manipulation of object motion and camera viewpoint while...

🔹 Publication Date: Published on Apr 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.07348
• PDF: https://arxiv.org/pdf/2604.07348
• Project Page: https://research.nvidia.com/labs/sil/projects/moright/

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Neural Computers

📝 Summary:
Neural Computers represent a new computing paradigm where models function as runtime systems, learning to execute tasks through I/O traces rather than explicit programming. AI-generated summary We pro...

🔹 Publication Date: Published on Apr 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.06425
• PDF: https://arxiv.org/pdf/2604.06425
• Project Page: https://metauto.ai/neuralcomputer/

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨RAGEN-2: Reasoning Collapse in Agentic RL

📝 Summary:
Research identifies template collapse in multi-turn LLM agents as a hidden failure mode undetectable by entropy, proposing mutual information proxies and SNR-aware filtering to improve reasoning quali...

🔹 Publication Date: Published on Apr 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.06268
• PDF: https://arxiv.org/pdf/2604.06268
• Project Page: https://ragen-ai.github.io/v2/
• Github: https://github.com/mll-lab-nu/RAGEN

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Improving Semantic Proximity in Information Retrieval through Cross-Lingual Alignment

📝 Summary:
Multilingual retrieval models exhibit bias toward English documents in mixed-language document pools, which is addressed through a novel training strategy that improves cross-lingual alignment with mi...

🔹 Publication Date: Published on Apr 7

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.05684
• PDF: https://arxiv.org/pdf/2604.05684

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨INSPATIO-WORLD: A Real-Time 4D World Simulator via Spatiotemporal Autoregressive Modeling

📝 Summary:
INSPATIO-WORLD presents a real-time framework for generating high-fidelity dynamic scenes from single videos using spatiotemporal autoregressive architecture and joint distribution matching distillati...

🔹 Publication Date: Published on Apr 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.07209
• PDF: https://arxiv.org/pdf/2604.07209
• Project Page: https://inspatio.github.io/inspatio-world/
• Github: https://github.com/inspatio/inspatio-world

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨VenusBench-Mobile: A Challenging and User-Centric Benchmark for Mobile GUI Agents with Capability Diagnostics

📝 Summary:
VenusBench-Mobile presents a comprehensive evaluation framework for mobile GUI agents that reveals significant performance gaps compared to existing benchmarks, emphasizing the need for more robust re...

🔹 Publication Date: Published on Feb 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.06182
• PDF: https://arxiv.org/pdf/2604.06182
• Github: https://github.com/inclusionAI/UI-Venus/tree/VenusBench-Mobile

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research