# Jan 9, 2025

## Daily

### NeuralSVG: An Implicit Representation for Text-to-Vector Generation
- **NeuralSVG** introduces an innovative approach to **text-to-vector graphics generation**, leveraging a small MLP network to encode entire scenes, enhancing the **layered structure** crucial for vector graphics.

### rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking
- **rStar-Math** demonstrates that **small language models (SLMs)** can achieve or exceed the math reasoning capabilities of larger models like OpenAI's o1 through innovative techniques such as **Monte Carlo Tree Search (MCTS)** and a self-evolution process.

### Nvidia releases its own brand of world models
- **Nvidia's Cosmos World Foundation Models (Cosmos WFMs)** are now available, enabling developers to create **physics-aware** videos and synthetic data for applications like robotics and autonomous vehicles, with models ranging from **4 billion to 14 billion parameters**.

### TabPFN v2: Accurate predictions on small data with a tabular foundation model
- **TabPFN v2** is a **pretrained transformer** that excels in small tabular data, achieving superior performance in **2.8 seconds for classification** and **4.8 seconds for regression** tasks, even outperforming strong baselines tuned for hours.

### Show HN: TabPFN v2 – A SOTA foundation model for small tabular data
- **TabPFN** is a novel tabular foundation model that significantly outperforms traditional methods, achieving superior predictions on datasets with up to **10,000 samples** in just **2.8 seconds**, compared to **4 hours** for conventional models like gradient-boosted decision trees.

### Video-of-Thought: Step-by-Step Video Reasoning from Perception to Cognition
- The **Video-of-Thought (VoT)** framework introduces a **novel Multimodal Large Language Model (MLLM)**, MotionEpic, which enhances video comprehension by integrating **spatial-temporal scene graph (STSG)** representation for pixel-level grounding.

### ObliqueTree: Advanced Decision Tree Implementation
- **ObliqueTree** offers a **high-performance** decision tree implementation that excels in both classification and regression tasks, utilizing **oblique splits** for enhanced flexibility and generalization with shallow trees.

### LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks
- **LongBench v2** is a comprehensive benchmark featuring **503 challenging questions** across six task categories, designed to evaluate LLMs' capabilities in handling long-context problems that require deep understanding and reasoning, with contexts ranging from **8k to 2M words**.

### Unveiling a New Era of Local AI With NVIDIA NIM Microservices and AI Blueprints
- **NVIDIA's new NIM microservices and AI Blueprints** enable generative AI on RTX AI PCs, enhancing capabilities in digital human creation, content generation, and productivity applications, all powered by the advanced GeForce RTX 50 Series GPUs.

### Hyundai Motor Group Embraces NVIDIA AI and Omniverse for Next-Gen Mobility
- **Hyundai Motor Group** is leveraging **NVIDIA AI** and **Omniverse** technologies to enhance vehicle safety, manufacturing efficiency, and robotics, marking a significant step towards **next-gen mobility** solutions.

### CO₂ Emissions and Models Performance: Insights from the Open LLM Leaderboard
- **CO₂ emissions from model inference** have become a critical concern, with recent evaluations revealing that community fine-tunes often outperform official models in carbon efficiency, highlighting a shift towards more sustainable AI practices.

### Supervision-free Vision-Language Alignment
- **SVP (Supervision-free Visual Projection)** enhances vision-language alignment by utilizing self-captioning and a pre-trained grounding model, eliminating the need for curated image-text pairs.

### CURing Large Models: Compression via CUR Decomposition
- **CURing** leverages **CUR matrix decomposition** to compress large models, approximating weight matrices with selected columns (C), rows (R), and a linking matrix (U), achieving significant size reduction with minimal performance loss.

### Integrating Ascend Backend with Torchtune through PyTorch Multi-Device Support
- **Torchtune** is a **PyTorch-native library** that simplifies the fine-tuning of Large Language Models (LLMs) by providing modular building blocks and supporting various training methods across different GPUs, including **Ascend NPU** integration for enhanced performance.