✨ Title: HyperClick: Advancing Reliable GUI Grounding via Uncertainty Calibration

📝 Summary:
GUI agents are overconfident and unreliable in grounding. HyperClick improves reliability by a dual reward mechanism that calibrates spatial confidence, reducing overconfidence. It achieves state-of-the-art performance for dependable GUI automation.

🔹 Publication Date: Published on Oct 31

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.27266
• PDF: https://arxiv.org/pdf/2510.27266
• Github: https://github.com/xiaomi-research/hyperclick

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

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

✨ Title: Defeating the Training-Inference Mismatch via FP16

📝 Summary:
RL fine-tuning of LLMs is unstable due to a numerical mismatch caused by BF16s rounding errors. We found that simply using FP16 effectively resolves this issue, leading to more stable optimization, faster convergence, and stronger performance. This simple change requires no model or algorithm mod...

🔹 Publication Date: Published on Oct 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.26788
• PDF: https://arxiv.org/pdf/2510.26788
• Github: https://github.com/sail-sg/Precision-RL

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

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

✨ Title: Phased DMD: Few-step Distribution Matching Distillation via Score Matching within Subintervals

📝 Summary:
To overcome limitations of one-step DMD on complex generative tasks, Phased DMD proposes a multi-step distillation framework. It employs progressive distribution matching across SNR subintervals with score matching to enhance diversity and generative capabilities.

🔹 Publication Date: Published on Oct 31

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.27684
• PDF: https://arxiv.org/pdf/2510.27684

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

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

✨ Title: Revisiting Multimodal Positional Encoding in Vision-Language Models

📝 Summary:
This paper systematically analyzes multimodal Rotary Positional Embedding RoPE for vision-language models. It identifies key guidelines for its design and proposes MHRoPE and MRoPE-Interleave, simple variants that significantly improve multimodal understanding.

🔹 Publication Date: Published on Oct 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.23095
• PDF: https://arxiv.org/pdf/2510.23095

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

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

✨ Title: Higher-order Linear Attention

📝 Summary:
Higher-order Linear Attention HLA addresses the quadratic cost of standard attention. It offers a scalable causal streaming mechanism for higher-order interactions with constant state size and linear per-token computation. HLA combines attention-like mixing with efficient recurrent architectures.

🔹 Publication Date: Published on Oct 31

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.27258
• PDF: https://arxiv.org/pdf/2510.27258
• Project Page: https://yifanzhang-pro.github.io/HLA
• Github: https://github.com/yifanzhang-pro/HLA

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

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

✨ Title: Dual-Stream Diffusion for World-Model Augmented Vision-Language-Action Model

📝 Summary:
DUST is a novel dual-stream diffusion framework for world-model augmented VLAs. It resolves modality conflicts by using separate streams for vision and action, enabling joint prediction without a unified latent space. DUST achieves significant performance gains in both simulation and real-world r...

🔹 Publication Date: Published on Oct 31

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.27607
• PDF: https://arxiv.org/pdf/2510.27607

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

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

✨ Title: The Denario project: Deep knowledge AI agents for scientific discovery

📝 Summary:
Denario is an AI multi-agent system for scientific research. It handles tasks like idea generation, code execution, and paper drafting. It successfully generated multiple scientific papers across diverse disciplines, expert-evaluated.

🔹 Publication Date: Published on Oct 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.26887
• PDF: https://arxiv.org/pdf/2510.26887
• Github: https://github.com/AstroPilot-AI/Denario

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

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

✨ Title: Visual Backdoor Attacks on MLLM Embodied Decision Making via Contrastive Trigger Learning

📝 Summary:
This paper introduces BEAT, the first framework for visual backdoor attacks on MLLM embodied agents using object triggers. It uses diverse training data and Contrastive Trigger Learning to ensure precise backdoor activation. BEAT achieves high attack success and exposes a critical security risk.

🔹 Publication Date: Published on Oct 31

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.27623
• PDF: https://arxiv.org/pdf/2510.27623
• Project Page: https://zqs1943.github.io/BEAT/

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

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

✨ Title: Mask-to-Height: A YOLOv11-Based Architecture for Joint Building Instance Segmentation and Height Classification from Satellite Imagery

📝 Summary:
This paper applies YOLOv11, a new deep learning model, for joint building instance segmentation and discrete height classification from satellite imagery. It achieves strong performance on the DFC2023 dataset, outperforming earlier models in accuracy and speed for urban mapping.

🔹 Publication Date: Published on Oct 31

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.27224
• PDF: https://arxiv.org/pdf/2510.27224

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

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

✨ Title: Limits of Generalization in RLVR: Two Case Studies in Mathematical Reasoning

📝 Summary:
This paper investigates RLVR for mathematical reasoning in LLMs using two combinatorial problems. It finds that while RLVR improves performance, it often reinforces superficial heuristics rather than genuine new reasoning strategies. This highlights RLVRs generalization limits and the need for be...

🔹 Publication Date: Published on Oct 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.27044
• PDF: https://arxiv.org/pdf/2510.27044
• Github: https://github.com/xashru/rlvr-seq-generalization

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

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

✨ Title: A Survey on Efficient Vision-Language-Action Models

📝 Summary:
This survey reviews Efficient Vision-Language-Action models Efficient VLAs, which address the high computational and data requirements of existing VLAs. It categorizes efficiency techniques into model design, training, and data collection, providing a comprehensive overview and future roadmap.

🔹 Publication Date: Published on Oct 27

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.24795
• PDF: https://arxiv.org/pdf/2510.24795
• Project Page: https://evla-survey.github.io/
• Github: https://github.com/YuZhaoshu/Efficient-VLAs-Survey

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

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

✨ Title: Value Drifts: Tracing Value Alignment During LLM Post-Training

📝 Summary:
This paper traces how LLM value alignment emerges during post-training, not just in final models. It finds supervised fine-tuning SFT primarily sets model values, with preference optimization rarely shifting them. Different preference optimization algorithms also yield varied alignment outcomes.

🔹 Publication Date: Published on Oct 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.26707
• PDF: https://arxiv.org/pdf/2510.26707

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

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

🤖🧠 HunyuanWorld-Mirror: Tencent’s Breakthrough in Universal 3D Reconstruction

🗓️ 03 Nov 2025
📚 AI News & Trends

The race toward achieving universal 3D understanding has reached a significant milestone with Tencent’s HunyuanWorld-Mirror, a cutting-edge open-source model designed to revolutionize 3D reconstruction. In an era dominated by visual intelligence and immersive digital experiences, this new model stands out by offering a feed-forward, geometry-aware framework that can predict multiple 3D outputs in a single ...

#HunyuanWorld #Tencent #3DReconstruction #UniversalAI #GeometryAware #OpenSourceAI

✨ Title: SemCoT: Accelerating Chain-of-Thought Reasoning through Semantically-Aligned Implicit Tokens

📝 Summary:
Chain-of-Thought CoT reasoning is verbose. SemCoT accelerates implicit CoT by ensuring semantic alignment of reasoning steps and speeding up individual implicit token generation. It uses a contrastive sentence transformer and an efficient, lightweight reasoning generator, outperforming state-of-t...

🔹 Publication Date: Published on Oct 28

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.24940
• PDF: https://arxiv.org/pdf/2510.24940
• Github: https://github.com/YinhanHe123/SemCoT/

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

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

✨ Title: Rank-GRPO: Training LLM-based Conversational Recommender Systems with Reinforcement Learning

📝 Summary:
ConvRec-R1, a two-stage framework, enhances LLM-based conversational recommender systems. It uses behavioral cloning for quality data and introduces Rank-GRPO, an RL method tailored for rank-style outputs. This improves recommendation quality, convergence, Recall, and NDCG.

🔹 Publication Date: Published on Oct 23

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.20150
• PDF: https://arxiv.org/pdf/2510.20150
• Github: https://github.com/yaochenzhu/Rank-GRPO

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

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

✨ Title: MisSynth: Improving MISSCI Logical Fallacies Classification with Synthetic Data

📝 Summary:
Misinformation is difficult to classify. MisSynth uses RAG to create synthetic fallacy data for LLM fine-tuning. This pipeline substantially improves LLM accuracy in identifying scientific misinformation fallacies, with over 35% F1-score gains.

🔹 Publication Date: Published on Oct 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.26345
• PDF: https://arxiv.org/pdf/2510.26345
• Github: https://github.com/mxpoliakov/MisSynth

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

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

✨ Title: Monopoly Deal: A Benchmark Environment for Bounded One-Sided Response Games

📝 Summary:
A new game structure, Bounded One-Sided Response Games BORGs, involves actions briefly transferring control to an opponent to satisfy a condition. A modified Monopoly Deal is used as a benchmark, and standard CFR effectively learns strategies.

🔹 Publication Date: Published on Oct 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.25080
• PDF: https://arxiv.org/pdf/2510.25080

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

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

✨ Title: Beyond Objects: Contextual Synthetic Data Generation for Fine-Grained Classification

📝 Summary:
BOB is a T2I model fine-tuning strategy for synthetic data generation in low-shot fine-grained classification. It extracts class-agnostic attributes to condition fine-tuning, then marginalizes them out during generation. This mitigates overfitting and achieves state-of-the-art results.

🔹 Publication Date: Published on Oct 28

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.24078
• PDF: https://arxiv.org/pdf/2510.24078
• Github: https://github.com/princetonvisualai/BeyondObjects

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

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

✨ Title: Every Activation Boosted: Scaling General Reasoner to 1 Trillion Open Language Foundation

📝 Summary:
Ling 2.0 introduces reasoning-oriented language models, scaling to 1 trillion parameters using sparse Mixture-of-Experts. It leverages activated computation to boost reasoning efficiency and capability up to 7-fold compared to dense models. This demonstrates sparse activation enables scalable, ef...

🔹 Publication Date: Published on Oct 25

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.22115
• PDF: https://arxiv.org/pdf/2510.22115

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

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

✨ Title: Towards Universal Video Retrieval: Generalizing Video Embedding via Synthesized Multimodal Pyramid Curriculum

📝 Summary:
This paper presents a co-designed framework for universal video retrieval. It introduces the UVRB benchmark, synthesizes multimodal data, and devises a Modality Pyramid curriculum for the General Video Embedder GVE. GVE achieves state-of-the-art zero-shot generalization, highlighting limitations ...

🔹 Publication Date: Published on Oct 31

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2510.27571
• PDF: https://arxiv.org/pdf/2510.27571
• Project Page: https://gzn00417.github.io/GVE/

🔹 Models citing this paper:
• https://huggingface.co/Alibaba-NLP/GVE-3B
• https://huggingface.co/Alibaba-NLP/GVE-7B

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

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

✨ Title: Generalizing Test-time Compute-optimal Scaling as an Optimizable Graph

📝 Summary:
This paper optimizes multi-LLM collaboration graphs for TTS, finding compute-optimal designs. It proposes Agent-REINFORCE, an LLM-agent framework using textual feedback to efficiently find them. Outperforms baselines, balancing accuracy and latency.

🔹 Publication Date: Published on Oct 29

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2511.00086
• PDF: https://arxiv.org/pdf/2511.00086

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

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