Transformers, Diffusion, and Quantum Computing

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 21 new subscribers since last edition!

Let's get into it 👇

Contents

1. Enabling Population-Level Parallelism in Tree-Based Genetic Programming for Comprehensive GPU Acceleration

2. FlowSpec: Continuous Pipelined Speculative Decoding for Efficient Distributed LLM Inference

3. MAPS: Advancing Multi-Modal Reasoning in Expert-Level Physical Science

4. RefTok: Reference-Based Tokenization for Video Generation

5. Less is Enough: Training-Free Video Diffusion Acceleration via Runtime-Adaptive Caching

6. Bootstrapping Grounded Chain-of-Thought in Multimodal LLMs for Data-Efficient Model Adaptation

7. Transformers Don't Need LayerNorm at Inference Time: Scaling LayerNorm Removal to GPT-2 XL and the Implications for Mechanistic Interpretability

8. StructTransform: A Scalable Attack Surface for Safety-Aligned Large Language Models

9. Gradient-Based Model Fingerprinting for LLM Similarity Detection and Family Classification

10. Batch-Max: Higher LLM Throughput using Larger Batch Sizes and KV Cache Compression

11. GPAS: Accelerating Convergence of LLM Pretraining via Gradient-Preserving Activation Scaling

12. DeSTA2.5-Audio: Toward General-Purpose Large Audio Language Model with Self-Generated Cross-Modal Alignment

13. MultiGen: Using Multimodal Generation in Simulation to Learn Multimodal Policies in Real

14. Large Language Model-Driven Closed-Loop UAV Operation with Semantic Observations

15. DexVLG: Dexterous Vision-Language-Grasp Model at Scale

Enabling Population-Level Parallelism in Tree-Based Genetic Programming for Comprehensive GPU Acceleration

Authors: Zhihong Wu, Lishuang Wang, Kebin Sun, Zhuozhao Li, Ran Cheng

Source and references: https://arxiv.org/abs/2501.17168v4

---

Introduction

This paper presents EvoGP, a high-performance framework that enables comprehensive GPU acceleration of Tree-based Genetic Programming (TGP) through population-level parallel execution.