# Mar 15, 2025

## Block Diffusion: Interpolating between autoregressive and diffusion models

- **Block diffusion language models** bridge the gap between autoregressive and diffusion models, enhancing **flexible-length generation** and inference efficiency through techniques like KV caching and parallel token sampling.

## Transformers Without Normalization

- **Dynamic Tanh (DyT)** serves as a **drop-in replacement** for normalization layers in Transformers, enabling models to achieve **comparable or superior performance** without the need for hyperparameter tuning.

## Ask HN: Any insider takes on Yann LeCun's push against current architectures?

- **Yann LeCun argues** that current LLM architectures will perpetually struggle with hallucinations due to **runaway errors** from token choice methods, suggesting a shift towards an **'energy minimization' architecture** that evaluates the overall response energy during training.

## Mayo Clinic's secret weapon against AI hallucinations: Reverse RAG in action

- **Mayo Clinic** employs a **reverse RAG** technique to combat AI hallucinations, effectively linking extracted data back to its original sources, which has nearly eradicated retrieval-based inaccuracies in non-diagnostic scenarios.

## [R] Transformers without Normalization (FAIR Meta, New York University, MIT, Princeton University)

- **Transformers without normalization** can achieve equal or superior performance through a novel technique called **Dynamic Tanh (DyT)**, which replaces traditional normalization layers with a simple element-wise operation, DyT(x)=tanh(αx).

## Arbitrary-Scale Super-Resolution with Neural Heat Fields

- **Thera** employs a **hypernetwork** to estimate pixel-wise parameters, ensuring **aliasing-free** super-resolution by utilizing local neural heat fields that adapt based on frequency and scaling factors.

## Show HN: OCR Benchmark Focusing on Automation

- **Document processing automation** is increasingly critical as new OCR solutions emerge, yet enterprises face challenges in discerning genuine capabilities from marketing claims, necessitating robust benchmarks for evaluation.

## [R] How Pickle Files Backdoor AI Models—And What You Can Do About It

- **Pickle files** can introduce vulnerabilities in AI models by allowing malicious code execution during the deserialization process, potentially compromising model integrity and security.

## Show HN: Open-Source MCP Server for Context and AI Tools

- The **JigsawStack MCP Server** enables AI models to access **live search**, **OCR**, and **structured data extraction** through a standardized interface, enhancing their capabilities beyond fixed context windows and outdated knowledge.

## Owl: Optimized Workforce Learning for multi-agent collaboration

- **OWL is a pioneering framework** for multi-agent collaboration in task automation, achieving a **58.18 average score on the GAIA benchmark**, ranking it **#1 among open-source frameworks**.

## [R] Block Diffusion: A Hybrid Language Model Combining Autoregressive and Diffusion Approaches for Flexible-Length Generation

- **Block Diffusion** innovatively combines **autoregressive** and **diffusion** models by processing text in flexible blocks, achieving a **9.37 perplexity** on C4 validation, which sets a new **SOTA** for diffusion language models.

## [R] Multi-View Video Generation via View-Invariant Motion Learning and Cross-View Consistent Translation

- **Reangle-A-Video** innovatively frames **4D video generation** as a **video-to-video translation** problem, enabling the synthesis of arbitrary camera viewpoints from a single input video while ensuring **temporal consistency**.

## AutoHete: An Automatic and Efficient Heterogeneous Training System for LLMs

- **AutoHete** is an innovative training system that enhances the efficiency of **heterogeneous training** for large language models (LLMs) by dynamically optimizing resource allocation based on hardware and training needs.