Diffusion, Vision-Language, and Robotic Grasping: Advances Across AI Frontiers

Latest research summaries in ML, Robotics, CV, NLP and AI

Welcome to today's edition of State of AI 👋 And a warm welcome to our 142 new subscribers since last edition!

This edition covers some fascinating developments in diffusion models, the integration of vision and language for AI agents, and novel approaches to robotic grasping. From pushing the boundaries of generative AI to enabling more intelligent and capable robots, these papers showcase the rapid progress happening across multiple AI research frontiers.

Here's what caught our attention:

• Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction - A novel theoretical framework that unifies and improves upon over 10 existing one-step diffusion distillation approaches, achieving new state-of-the-art performance on image generation benchmarks.

• EgoVLA: Learning Vision-Language-Action Models from Egocentric Human Videos - A system that learns dexterous manipulation skills by leveraging the scale and diversity of human egocentric videos, enabling efficient transfer to robotic platforms.

• GraspGen: A Diffusion-based Framework for 6-DOF Grasping with On-Generator Training - A flexible, high-performing grasping framework that uses diffusion models and a novel on-generator training approach to recognize and filter out its own failure modes.

Let's get into it 👇

Contents

1. Voxtral

2. Towards Formal Verification of LLM-Generated Code from Natural Language Prompts

3. Prompt Injection 2.0: Hybrid AI Threats

4. Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models

5. Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction

6. GLAD: Generalizable Tuning for Vision-Language Models

7. Training Transformers with Enforced Lipschitz Constants

8. GeoReg: Weight-Constrained Few-Shot Regression for Socio-Economic Estimation using LLM

9. Federated Learning: A Survey on Privacy-Preserving Collaborative Intelligence

10. Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

11. MERA Code: A Unified Framework for Evaluating Code Generation Across Tasks

12. QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation

13. EgoVLA: Learning Vision-Language-Action Models from Egocentric Human Videos

14. DINO-VO: A Feature-based Visual Odometry Leveraging a Visual Foundation Model

15. GraspGen: A Diffusion-based Framework for 6-DOF Grasping with On-Generator Training

Voxtral

Source and references: https://arxiv.org/abs/2507.13264v1

---

Introduction

This paper presents Voxtral Mini and Voxtral Small, two multimodal audio chat models that can comprehend both spoken audio and text documents, achieving state-of-the-art performance across a diverse range of audio benchmarks while preserving strong text capabilities.

Key Points