From Bandit-Prompt Tuning to VTBench Visual Tokenizers and Greedy Decoding: Precision Tools for Scaling LLM Performance

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

🧠 Stay Focused, One Task at a Time — 30% Off Forget.work (Forever)

At State of AI, we’re always testing tools that make deep work easier. That’s why we’re excited to partner with Forget.work the #1 productivity tool designed for focused, distraction-free work.

Forget.work is purpose-built for single-tasking. It helps you beat time blindness and stay anchored to what you’re doing right now. Always visible on your screen, it gently reminds you to stay on task, no context switching, no endless tab juggling.

Whether you’re deep in a research paper, debugging an experiment, or trying to write without flipping to Slack or Twitter, Forget keeps your mind tethered to the task at hand.

As a State of AI reader, you get 30% off for life.
👉 Use the promo code STATEAI when checking out

10,000+ professionals already use it to get things done, ADHD or not. It’s like a subtle co-pilot for your workday, and we think you’ll love it.

---

Welcome to the 62 new subscribers since our last edition! Let’s jump straight into our AI research roundup!

Contents

1. AutoMathKG: The automated mathematical knowledge graph based on LLM and vector database

2. Robin: A multi-agent system for automating scientific discovery

3. Multi-Armed Bandits Meet Large Language Models

4. Mean Flows for One-step Generative Modeling

5. VTBench: Evaluating Visual Tokenizers for Autoregressive Image Generation

6. G1: Bootstrapping Perception and Reasoning Abilities of Vision-Language Model via Reinforcement Learning

7. Greed is Good: A Unifying Perspective on Guided Generation

8. Restoration Score Distillation: From Corrupted Diffusion Pretraining to One-Step High-Quality Generation

9. Occult: Optimizing Collaborative Communication across Experts for Accelerated Parallel MoE Training and Inference

10. Granary: Speech Recognition and Translation Dataset in 25 European Languages

11. SMOTExT: SMOTE meets Large Language Models

12. CIE: Controlling Language Model Text Generations Using Continuous Signals

13. Cosmos-Reason1: From Physical Common Sense To Embodied Reasoning

14. Seeing, Saying, Solving: An LLM-to-TL Framework for Cooperative Robots

15. GraspMolmo: Generalizable Task-Oriented Grasping via Large-Scale Synthetic Data Generation

AutoMathKG: The automated mathematical knowledge graph based on LLM and vector database

Authors: Rong Bian, Yu Geng, Zijian Yang, Bing Cheng

Source and references: https://arxiv.org/abs/2505.13406v1

---

Introduction

This paper proposes AutoMathKG, a high-quality, wide-coverage, and multi-dimensional knowledge graph (KG) for the field of mathematics, written in natural language and capable of automatic updates.