Large Language Models News and Research

RSS

MULTITuDE: A Benchmark for the Detection of Machine-Generated Text

This paper presents MULTITuDE, a benchmark dataset designed for multilingual machine-generated text (MGT) detection. The study evaluates various detection methods across 11 languages, demonstrating that fine-tuning detectors with multilingual language models is an effective approach, and the linguistic similarity between languages plays a significant role in the generalization of detectors.

26 Oct 2023

MarineGPT: Advancing Marine Vision-Language AI

MarineGPT, a groundbreaking vision-language model designed specifically for the marine domain, has been developed to identify marine objects from visual inputs and provide comprehensive, scientific, and sensitive responses. This model leverages the Marine-5M dataset and offers improved marine vision and language alignment, contributing to increased public awareness of marine biodiversity while addressing some limitations.

25 Oct 2023

Quality Diversity through Human Feedback (QDHF)

Researchers introduce a novel approach called Quality Diversity through Human Feedback (QDHF), which leverages human judgments to derive diversity metrics in Quality Diversity (QD) algorithms. This method, based on latent space projection and contrastive learning, offers a more adaptable and effective way to measure diversity, particularly in complex and abstract domains.

25 Oct 2023

Responsible Generative AI: Guidelines for Safeguarding Research Integrity

Researchers discussed the development of "Living guidelines for responsible use of generative artificial intelligence (AI) in research." These guidelines, crafted by a collaboration of international scientific institutions, organizations, and policy advisers, aim to address the potential risks posed by generative AI and provide key principles for its responsible use in scientific research.

25 Oct 2023

FACTCHD: Detecting Fact-Conflicting Hallucinations in Large Language Models

Researchers have introduced FACTCHD, a framework for detecting fact-conflicting hallucinations in large language models (LLMs). They developed a benchmark that provides interpretable data for evaluating the factual accuracy of LLM-generated responses and introduced the TRUTH-TRIANGULATOR framework to enhance hallucination detection.

23 Oct 2023

Automatic LLM Evaluation on Protocol Planning Using BioPlanner

Researchers introduced a framework, BioPlanner, for automatically evaluating the performance of large language models (LLMs) in planning experimental protocols in biology. They addressed the challenge of evaluating LLM-generated protocol accuracy by utilizing pseudocode representations and a dataset called BIOPROT, demonstrating the effectiveness of the approach in generating precise and executable laboratory protocols.

20 Oct 2023

Interactive Task Planning with Language Models

Researchers introduced an interactive robot framework that leverages Large Language Models (LLMs) to excel in long-term task planning, adapting to new goals and tasks during execution. The system seamlessly integrates high-level planning and low-level execution using language, demonstrating robustness and adaptability across tasks.

19 Oct 2023

AI in Scientific Publishing: Promise and Peril

The integration of generative artificial intelligence (GAI) in scientific publishing, exemplified by AI tools like ChatGPT and GPT-4, is transforming research paper writing and dissemination. While AI offers benefits such as expediting manuscript creation and improving accessibility, it raises concerns about inaccuracies, ethical considerations, and challenges in distinguishing AI-generated content.

12 Oct 2023

Neural Policy Gradient (PG) RANK: A Breakthrough in Training Language Models for Text Retrieval

Researchers introduced a novel training algorithm, Neural Policy Gradient (PG) RANK, which leverages policy gradient to train retrieval models end-to-end. This approach reduces reliance on heuristics and aligns the training objective with downstream decision-making quality, resulting in significant in-domain performance improvements and robust out-of-domain generalization in question-answering tasks.

11 Oct 2023

MathCoder: Advancing Mathematical Reasoning with Language Models

Researchers introduce MathCoder, an open-source language model fine-tuned for mathematical reasoning. MathCoder achieves state-of-the-art performance among open-source models, emphasizing the integration of reasoning, code generation, and execution. However, it faces challenges with complex geometry and theorem-proving problems, leaving room for future improvements.

10 Oct 2023

LanguageMPC: Large Language Models as Decision Makers for Autonomous Driving

This research explores the application of Large Language Models (LLMs) as decision-making components in autonomous driving (AD) systems, addressing challenges in understanding complex driving scenarios. The LLMs, equipped with reasoning skills, enhance the AD system's adaptability and transparency, effectively handling intricate driving situations, and offering a promising direction for future developments in this field.

9 Oct 2023

MindShift: A Dynamic Intervention for Problematic Smartphone Use

Researchers introduce MindShift, an innovative intervention technique driven by large language models (LLMs), to address problematic smartphone use by adapting persuasive content in real-time based on users' contexts and mental states. In a 5-week field experiment with 25 participants, MindShift significantly increased intervention acceptance rates and reduced smartphone use frequency.

4 Oct 2023

Disinformation Detection in the Age of LLMs: Challenges and Solutions

This research investigates the challenges of detecting misinformation generated by Large Language Models (LLMs) like ChatGPT. Existing detection techniques face difficulties in distinguishing LLM-generated disinformation, prompting the development of advanced prompt engineering methods to improve detection accuracy and counter the spread of misleading content.

2 Oct 2023

OmniEvent: A Comprehensive Toolkit for Event Understanding in Text

OmniEvent is an innovative toolkit for event understanding in text, addressing event detection (ED), event argument extraction (EAE), and event relation extraction (ERE). It offers a comprehensive, fair, and user-friendly approach, supporting various mainstream models and datasets while providing solutions to common evaluation pitfalls, making it a valuable tool in natural language processing research.

28 Sep 2023

REHEARSAL: Enhancing Conflict Resolution Skills through Simulated Interactions

Researchers introduce the REHEARSAL system, a novel approach to improving conflict resolution skills. This interactive dialogue system allows users to engage in realistic conflict simulations, receive valuable feedback, and explore alternative approaches, all guided by Interest-Rights-Power (IRP) prompting rooted in conflict resolution theory.

27 Sep 2023

Safeguarding Anonymized Data in the Age of Language Models: Challenges and Solutions

Researchers investigate the risks posed by Large Language Models (LLMs) in re-identifying individuals from anonymized texts. Their experiments reveal that LLMs, such as GPT-3.5, can effectively deanonymize data, raising significant privacy concerns and highlighting the need for improved anonymization techniques and privacy protection strategies in the era of advanced AI.

27 Sep 2023

ForceSight: Robust Mobile Manipulation Guided by Visual-Force Goals and Natural Language

Researchers at Georgia Tech have developed ForceSight, a system that combines deep neural networks and natural language instructions to enable robots to perform robust mobile manipulation tasks. ForceSight predicts visual-force goals and utilizes a vision transformer architecture to improve the accuracy of these predictions, significantly enhancing the robot's performance in various manipulation tasks.

26 Sep 2023

Language as the Medium: Text-Based Multimodal Video Classification

Researchers introduce a groundbreaking approach to multimodal classification using textual cues. Leveraging large language models (LLMs), this method interprets visual and auditory information through text, showcasing its effectiveness in video classification and offering opportunities for more complex contextual reasoning in future applications.

24 Sep 2023

Introducing Code Llama: Powerful Language Models for Efficient Coding

The Code Llama family of Large Language Models (LLMs) presents a breakthrough in coding capabilities. These models excel at code infilling and comprehensive input support, making them efficient coders without the need for prior examples. With three model types, including Python-specific variations, they offer a versatile toolkit for various coding tasks, underpinned by unique training methods and extensive datasets.

24 Sep 2023