# ML Times

## R-Zero: Self-Evolving Reasoning LLM from Zero Data  
**R-Zero** introduces a **fully autonomous framework** for training Large Language Models (LLMs) that generates its own data, eliminating reliance on human-curated tasks and labels, thus paving the way for **super-intelligence**.

## Defeating Nondeterminism in LLM Inference  
**Nondeterminism in LLM inference arises from floating-point non-associativity and batch size variability**, which can lead to different outputs even with identical inputs, as demonstrated by the varying completions generated by models like Qwen-3.

## Implementation and ablation study of the Hierarchical Reasoning Model (HRM): what really drives performance?  
The **Hierarchical Reasoning Model (HRM)** excels in performance primarily through **outer-loop refinement** with more segments, rather than its architecture, indicating a focus on training methodology over structural complexity.

## LLMs play a cooperative card game, coordination without communication  
**LLMs struggle with cooperative gameplay**, particularly in the card game _The Crew_, where smaller models fail to understand their roles, often prioritizing individual success over team strategy.

## News: Arm announces next Generation core family called Arm Lumex  
**Arm's Lumex platform** introduces **C1 CPUs** with **Scalable Matrix Extensions 2 (SME2)** and the **Mali G1-Ultra GPU**, designed specifically for AI applications in next-gen PCs and smartphones.

## UGMM-NN: Univariate Gaussian Mixture Model Neural Network  
The **uGMM-NN** introduces a **novel neural architecture** that embeds **probabilistic reasoning** into deep networks, allowing each node to parameterize activations as a **univariate Gaussian mixture** with learnable parameters.

## Interesting PEZY-SC4s  
**PEZY-SC4S** aims for **highly efficient FP64 compute** by utilizing a smaller die and lower power draw, achieving a projected **~91 GF/W** performance, significantly outperforming Nvidia's H200 and approaching AMD's MI300A.

## Graphrag pipeline that runs entirely locally with ollama and has full source attribution  
**VeritasGraph** is a locally-run Graph RAG pipeline utilizing **Ollama** with **Llama 3.1**, designed for private use and ensuring **full source attribution** for generated content.

## NVIDIA Blackwell Ultra crushes MLPerf  
**NVIDIA's Blackwell Ultra** achieved **5× throughput** on DeepSeek-R1 and set records on Llama 3.1 and Whisper, showcasing innovative techniques like **FP8 KV-cache** and **disaggregated serving**.

## Stability AI Introduces Stable Audio 2.5, the First Audio Model Built for Enterprise Sound Production at Scale  
**Stable Audio 2.5** is the first audio generation model tailored for **enterprise-grade sound production**, enabling brands to create distinct audio identities across various channels, enhancing memorability by up to **eight times**.

## Jupyter Agents: training LLMs to reason with notebooks  
**Jupyter Agents** empower LLMs to execute code within Jupyter Notebooks, enhancing their ability to tackle complex data analysis tasks through direct code execution and reasoning.

## NVIDIA Blackwell Ultra Sets the Bar in New MLPerf Inference Benchmark  
The **NVIDIA GB300 NVL72** rack-scale system achieves **record-breaking throughput** on the new reasoning inference benchmark, outperforming previous models by **up to 1.4x** in MLPerf Inference v5.1.

## NVIDIA Partners With AI Infrastructure Ecosystem to Unveil Reference Design for Giga-Scale AI Factories  
**NVIDIA's new reference design aims to revolutionize data centers into integrated AI factories**, enhancing energy efficiency and performance through collaboration with industry partners like Jacobs and Siemens Energy.

## Breaking the networking wall in AI infrastructure  
**MOSAIC technology** aims to resolve the **power, reliability, and reach trade-off** in AI infrastructure by utilizing a novel optical link design that combines low power consumption with high reliability and long distances, achieving up to **50 meters** of reach.