Bi-Weekly AI Research Roundup

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

Contents

1. LLM Echo Chamber: personalized and automated disinformation

2. DataGpt-SQL-7B: An Open-Source Language Model for Text-to-SQL

3. Leveraging Mixture of Experts for Improved Speech Deepfake Detection

4. OmniBench: Towards The Future of Universal Omni-Language Models

5. MonoFormer: One Transformer for Both Diffusion and Autoregression

6. CDChat: A Large Multimodal Model for Remote Sensing Change Description

7. Order of Magnitude Speedups for LLM Membership Inference

8. PICL: Physics Informed Contrastive Learning for Partial Differential Equations

9. Learning To Help: Training Models to Assist Legacy Devices

10. EuroLLM: Multilingual Language Models for Europe

11. DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs

12. MINERS: Multilingual Language Models as Semantic Retrievers

13. SceneVerse: Scaling 3D Vision-Language Learning for Grounded Scene Understanding

14. Will Large Language Models be a Panacea to Autonomous Driving?

15. MHRC: Closed-loop Decentralized Multi-Heterogeneous Robot Collaboration with Large Language Models

---

LLM Echo Chamber: personalized and automated disinformation

Authors: Tony Ma

Source and references: https://arxiv.org/abs/2409.16241v1

---

Introduction

This paper explores the risks associated with Large Language Models (LLMs) and their potential to disseminate misinformation. It presents the development of the "LLM Echo Chamber", a controlled digital environment designed to simulate the dynamics of social media platforms where misinformation often proliferates.

Key Points

• Identification of LLMs and their Vulnerabilities: Analyzing current state-of-the-art LLMs and identifying inherent vulnerabilities that could be exploited for misinformation generation.

• Development of an "LLM Echo Chamber": Creating a controlled environment that simulates the spread of misinformation by LLMs, allowing for the observation of user interactions.

• Persuasiveness and harmfulness analysis: Demonstrating the potential harmfulness and persuasiveness of LLM-generated misinformation, underscoring the critical need for ethical considerations and the development of robust strategies to counter these risks.

Methodology

The researchers systematically evaluated a set of leading LLMs, including GPT-3.5, Llama2, Phi2, and Gemma, to understand their capabilities and limitations. They then conducted finetuning experiments on these models using an identity-shifting dataset to jailbreak the models and generate misinformation. Finally, they developed the "LLM Echo Chamber" using Streamlit and LangChain, leveraging the finetuned Phi2 model to populate the chatroom with persuasive misinformation.

Results and Findings

The automated evaluation of the "LLM Echo Chamber" interactions using GPT-4 revealed that the misinformation generated by the finetuned LLMs was highly harmful (average score of 4.21 out of 5) and persuasive (average score of 3.24 out of 5). These results demonstrate the potential for LLMs to shape public opinion and discourse through the dissemination of misinformation.

Implications and Conclusions

The findings of this study underscore the urgent need for robust mechanisms to mitigate the risks associated with the use of LLMs. The ability of these models to generate persuasive and contextually relevant misinformation that can reinforce echo chambers and influence public discourse poses significant challenges for maintaining information integrity and reliability online. The researchers emphasize the importance of ongoing vigilance, the formulation of ethical guidelines, and the development of countermeasures to address the potential harms of LLM-driven misinformation.

---

DataGpt-SQL-7B: An Open-Source Language Model for Text-to-SQL

Authors: Lixia Wu, Peng Li, Junhong Lou, Lei Fu

Source and references: https://arxiv.org/abs/2409.15985v1

---

Introduction

This paper presents DataGPT-SQL-7B, a novel open-source language model for translating natural language queries into Structured Query Language (SQL) commands. The proposed model aims to democratize data access and analysis for non-expert users, mitigating risks associated with closed-source Large Language Models (LLMs).