AI Research Roundup: Diffusion Autoencoders, LLM Safeguards, Delta Compression, Neural Rendering & Automated Pen Testing 🔬

Breaking Down This Week's Most Impactful Papers in AI: From Novel Compression Methods to Secure LLM Architecture

Contents

1. Diffusion Autoencoders are Scalable Image Tokenizers

2. Advances in Multimodal Adaptation and Generalization: From Traditional Approaches to Foundation Models

3. Towards Automated Penetration Testing: Introducing LLM Benchmark, Analysis, and Improvements

4. Diffusion Autoencoders are Scalable Image Tokenizers

5. CLoQ: Enhancing Fine-Tuning of Quantized LLMs via Calibrated LoRA Initialization

6. More Expressive Attention with Negative Weights

7. Advances in Multimodal Adaptation and Generalization: From Traditional Approaches to Foundation Models

8. DiffusionRenderer: Neural Inverse and Forward Rendering with Video Diffusion Models

9. Inkspire: Supporting Design Exploration with Generative AI through Analogical Sketching

10. Leveraging LLM Agents for Automated Optimization Modeling for SASP Problems: A Graph-RAG based Approach

11. Automated Test-Case Generation for REST APIs Using Model Inference Search Heuristic

12. GuardReasoner: Towards Reasoning-based LLM Safeguards

13. DeltaLLM: Compress LLMs with Low-Rank Deltas between Shared Weights

14. Streaming DiLoCo with overlapping communication: Towards a Distributed Free Lunch

15. LLM-AutoDiff: Auto-Differentiate Any LLM Workflow

Diffusion Autoencoders are Scalable Image Tokenizers

Authors: Yinbo Chen, Rohit Girdhar, Xiaolong Wang, Sai Saketh Rambhatla, Ishan Misra

Source and references: https://arxiv.org/abs/2501.18593v1

---

Introduction

This paper presents a simple diffusion tokenizer (DiTo) that learns compact visual representations for image generation models, using a single diffusion L2 loss as the training objective.