Limitations of Large Language Models

Ограничения больших языковых моделей
Erin Sanu, T Keerthi Amudaa, Prasiddha Bhat, Guduru Dinesh, Apoorva Uday Kumar Chate, Ramakanth Kumar P
2024-11-07

Adversarial attacksAlgorithmic biasDomain-specific customizationLarge language modelsModel hallucinations
Large Language Models (LLMs) have become a cornerstone of modern natural language processing, exhibiting remarkable capabilities in diverse applications. However, these models are not flawless. This paper provides a comprehensive analysis of the loopholes inherent in LLMs, focusing on adversarial attacks, biases, hallucinations and outdatedness. One of the main problems is hallucinations, where models generate plausible but inaccurate or unreal information. Biases introduced into training data can lead to outcomes that reflect and amplify societal prejudices leading to discriminatory behavior. Adversarial attacks exploit vulnerabilities in machine learning models by introducing subtle perturbations to input data, causing the model to make incorrect predictions with high confidence. Additionally, LLMs considerably struggle when it comes to domain-specific queries as they lack specialized knowledge and have a generalized training, thus highlighting the need for domain-specific customization. By inspecting these vulnerabilities, we aim to highlight the critical areas that require attention to ensure the safe and effective deployment of LLMs.
1
Biases in training data may be reproduced and amplified by LLMs, resulting in discriminatory behavior.
2
LLM hallucinations can produce plausible but inaccurate or entirely fabricated information.
3
LLMs struggle with domain-specific queries because generalized training provides insufficient specialized knowledge, motivating domain-specific customization.
4
Subtle adversarial input perturbations can exploit model vulnerabilities and induce incorrect, highly confident predictions.
5
The paper identifies hallucinations, biases, adversarial attacks, and outdated knowledge as major limitations affecting large language models.

Large Language Models (LLMs)

Limitations and vulnerabilities of LLMs, including adversarial attacks, biases, hallucinations, outdatedness, and inadequate performance on domain-specific queries

Publication Details
Publication Date
2024-11-07
Journal
Publisher
ISSN
Cited by
13
Access Type
Author Information
Authors
Erin Sanu
T Keerthi Amudaa
Prasiddha Bhat
Guduru Dinesh
Apoorva Uday Kumar Chate
Ramakanth Kumar P
Explore further
Open the scid.ai AI chat with a ready-made request: it will find papers on a similar topic and help build a literature review.
Find similar papers in the chat →
Make a presentation
100%