# Jan 4, 2026

## Scaling Latent Reasoning via Looped Language Models
- **Ouro** introduces **Looped Language Models (LoopLM)** that integrate reasoning into the **pre-training phase**, enhancing performance without merely increasing model size, achieving results comparable to **12B SOTA LLMs** across various benchmarks.

## Recursive Language Models
- **Recursive Language Models (RLMs)** enable large language models (LLMs) to process prompts significantly longer than their typical context windows by allowing them to **decompose** and **recursively call** themselves on prompt snippets, enhancing their performance.

## Sirius DB
- **Sirius is a GPU-native SQL engine** that integrates seamlessly with existing databases like DuckDB, achieving over **10× speedup** in query execution without requiring major system changes.

## Interactive visualization of DeepSeek's mHC - why doubly stochastic constraints fix Hyper-Connection instability
- **Doubly stochastic constraints** effectively stabilize Hyper-Connection instability in DeepSeek's mHC by ensuring that matrix multiplications remain bounded, preventing exponential growth in deep networks.

## Anatomy of BoltzGen
- **BoltzGen** is a **Hugging Face Space** that showcases innovative applications of generative models, emphasizing the intersection of **machine learning** and **creativity**.

## Built a US Mortgage Underwriting OCR System With 96% Real-World Accuracy → Saved ~$2M Per Year
- A **US mortgage underwriting OCR system** achieves **~96% accuracy** in real-world applications, significantly outperforming typical services that plateau around **70–72%** accuracy, leading to substantial operational improvements.

## Dynamic Large Concept Models: Latent Reasoning in an Adaptive Semantic Space