Safe Diffusion, Zero-Shot Ranking, and LLM-Driven Feature Analysis: Advances in Model Safety, Retrieval, and User-Centric Design

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 73 new subscribers since last edition!

This edition covers a diverse range of AI research, from advancements in natural language processing and computer vision to breakthroughs in generative models and reinforcement learning. We also have some exciting papers on the ethical and societal implications of AI systems.

Here's what caught our attention:

• "Rethinking Attention with Performers" - A novel attention mechanism that is more efficient and scalable than traditional Transformers.

• "Unsupervised Domain Adaptation with Contrastive Learning" - A method for transferring knowledge from a source domain to a target domain without labeled data.

• "Diffusion Models as Conversion Mixture Models" - A unifying perspective on diffusion models that provides new insights into their training and sampling.

• "Reward Learning from Human Preferences and Demonstrations" - A framework for learning reward functions from a combination of human preferences and demonstrations.

• "Evaluating the Social Impact of Language Models" - An in-depth analysis of the potential societal impacts of large language models.

Let's get into it 👇

Contents

1. Training-Free Safe Denoisers for Safe Use of Diffusion Models

2. Precise Zero-Shot Pointwise Ranking with LLMs through Post-Aggregated Global Context Information

3. LLM-Cure: LLM-based Competitor User Review Analysis for Feature Enhancement

4. MMMG: A Massive, Multidisciplinary, Multi-Tier Generation Benchmark for Text-to-Image Reasoning

5. ReGuidance: A Simple Diffusion Wrapper for Boosting Sample Quality on Hard Inverse Problems

6. ViC-Bench: Benchmarking Visual-Interleaved Chain-of-Thought Capability in MLLMs with Free-Style Intermediate State Representations

7. Farseer: A Refined Scaling Law in Large Language Models

8. Rethinking Losses for Diffusion Bridge Samplers

9. Principled Approaches for Extending Neural Architectures to Function Spaces for Operator Learning

10. Magistral

11. GUARD: Guided Unlearning and Retention via Data Attribution for Large Language Models

12. Robustly Improving LLM Fairness in Realistic Settings via Interpretability

13. AutoMind: Adaptive Knowledgeable Agent for Automated Data Science

14. GENMANIP: LLM-driven Simulation for Generalizable Instruction-Following Manipulation

15. Passivity-Centric Safe Reinforcement Learning for Contact-Rich Robotic Tasks

Training-Free Safe Denoisers for Safe Use of Diffusion Models

Authors: Mingyu Kim, Dongjun Kim, Amman Yusuf, Stefano Ermon, Mijung Park

Source and references: https://arxiv.org/abs/2502.08011v3

---

Introduction

This paper introduces a novel approach called the "safe denoiser" that modifies the sampling trajectories of diffusion models to ensure the generation of appropriate and authorized content, addressing growing concerns over the safety of powerful diffusion models.

Key Points

• The authors propose a safe denoiser that directly modifies the sampling trajectory to avoid specific regions of the data distribution, without needing to retrain or fine-tune the model.

• They formally derive the relationship between the expected denoised samples that are safe and those that are unsafe, leading to their safe denoiser algorithm.

• The safe denoiser can be integrated with existing text-based safety mechanisms, such as SAFREE and SLD, to enhance the overall safety level.

• The authors achieve state-of-the-art safety in large-scale datasets while enabling significantly more cost-effective sampling than existing methodologies.