Decentralized LLM Serving, Trustworthy Decision Support, and Interpretable Sparse Autoencoders

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

Welcome to today’s edition of State of AI 🤖 👋

This edition covers decentralized systems for democratizing large language model serving, and trustworthy real-time decision support using low-latency interpretable models, all the way to novel autoencoder architectures that improve model interpretability and steerability.

Here’s what caught our attention:

• PlanetServe: A Decentralized, Scalable, and Privacy-Preserving Overlay for Democratizing Large Language Model Serving - A novel decentralized serving infrastructure that aims to make it easier for small organizations and individuals to deploy and experiment with their own LLM innovations.

• Achieving Trustworthy Real-Time Decision Support Systems with Low-Latency Interpretable AI Models - An in-depth review of the state-of-the-art in human-AI collaboration for decision-making, highlighting the need for more effective and generalizable approaches.

• Interpretable and Steerable Concept Bottleneck Sparse Autoencoders - A framework that combines the unsupervised discovery capabilities of sparse autoencoders with the controllability of concept bottlenecks, significantly improving model interpretability and steerability.

Let’s get into it 👇

Contents

1. PlanetServe: A Decentralized, Scalable, and Privacy-Preserving Overlay for Democratizing Large Language Model Serving

2. EcomBench: Towards Holistic Evaluation of Foundation Agents in E-commerce

3. Achieving Trustworthy Real-Time Decision Support Systems with Low-Latency Interpretable AI Models

4. Graph Laplacian Transformer with Progressive Sampling for Prostate Cancer Grading

5. Interpretable and Steerable Concept Bottleneck Sparse Autoencoders

6. Equivariant Test-Time Training with Operator Sketching for Imaging Inverse Problems

7. Scaling Behavior of Discrete Diffusion Language Models

8. SparseSwaps: Tractable LLM Pruning Mask Refinement at Scale

9. LLM-Driven Composite Neural Architecture Search for Multi-Source RL State Encoding

10. LabelFusion: Learning to Fuse LLMs and Transformer Classifiers for Robust Text Classification

11. The FACTS Leaderboard: A Comprehensive Benchmark for Large Language Model Factuality

12. Better Language Model Inversion by Compactly Representing Next-Token Distributions

13. MaskedManipulator: Versatile Whole-Body Manipulation

14. Reframing Human-Robot Interaction Through Extended Reality: Unlocking Safer, Smarter, and More Empathic Interactions with Virtual Robots and Foundation Models

15. Visualization Generation with Large Language Models: An Evaluation

PlanetServe: A Decentralized, Scalable, and Privacy-Preserving Overlay for Democratizing Large Language Model Serving