Author:
(1) Andrew J. Peterson, University of Poitiers (andrew.peterson@univ-poitiers.fr). Table of Links Abstract and Introduction Related Work The media, filter bubbles and echo chambers Network effects and Information Cascades Model collapse Known biases in LLMs A Model of Knowledge Collapse Results Discussion and References Appendix Comparing width of the tails Defining knowledge collapse Known biases in LLMs Newer AI models such as LLMs are not immune to the problems of bias identified and measured in machine learning algorithms (Nazer et al., 2023) and which have plagued predictive algorithms in real-world uses cases going back to at least the 1930s (Christian, 2021, Ch.2). Unsurprisingly, LLMs are better at recalling facts that occur frequently within the training data and struggle with long-tail knowledge (Kandpal et al., 2023). Das et al. (2024) identify a range of shortcomings of LLMs in attempting to generate human-like texts, such as underrepresenting minority viewpoints and reducing the broad concept of “positive” text to that simply of expressing “joy”. Recent work attempts to address these issues through a variety of methods, for example by upsampling underrepresented features on which prediction is otherwise sub-optimal (Gesi et al., 2023), or evaluating the importance of input data using shapely values (Karlas et al ˇ ., 2022). However, the mechanistic interpretability work on LLMs to date suggest that our understanding, while improving, is still very limited (e.g. Kramar et al ´ ., 2024; Wu et al., 2023). As such, direct methods for overcoming such biases are, at a minimum, not close at hand. Finally, while much of the focus is naturally on overt racial and gender biases, there may also be pervasive but less observable biases in the content and form of the output. Wendler et al. (2024), for example, provide evidence that that current LLMs trained on large amounts of English text ‘rely on’ English in their latent representations, as if a kind of reference language. One particular area in which the diversity of LLM outputs has been analyzed is on a token-by-token level in the context of decoding strategies. In some situations, using beam search to choose the most likely next token can create degenerate repetitive phrases (Su et al., 2022). Furthermore, a bit like Thelonious Monk’s melodic lines, humans do not string together sequences of the most likely words but occasionally try to surprise the listener by sampling from low-probability words, defying conventions, etc. Holtzman et al. (2020) (referring to Grice, 1975). This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. Author: (1) Andrew J. Peterson, University of Poitiers (andrew.peterson@univ-poitiers.fr). Author: Author: (1) Andrew J. Peterson, University of Poitiers (andrew.peterson@univ-poitiers.fr). Table of Links Abstract and Introduction Abstract and Introduction Related Work Related Work The media, filter bubbles and echo chambers The media, filter bubbles and echo chambers Network effects and Information Cascades Network effects and Information Cascades Model collapse Model collapse Known biases in LLMs Known biases in LLMs A Model of Knowledge Collapse A Model of Knowledge Collapse Results Results Discussion and References Discussion and References Appendix Appendix Appendix Comparing width of the tails Comparing width of the tails Defining knowledge collapse Defining knowledge collapse Known biases in LLMs Newer AI models such as LLMs are not immune to the problems of bias identified and measured in machine learning algorithms (Nazer et al., 2023) and which have plagued predictive algorithms in real-world uses cases going back to at least the 1930s (Christian, 2021, Ch.2). Unsurprisingly, LLMs are better at recalling facts that occur frequently within the training data and struggle with long-tail knowledge (Kandpal et al., 2023). Das et al. (2024) identify a range of shortcomings of LLMs in attempting to generate human-like texts, such as underrepresenting minority viewpoints and reducing the broad concept of “positive” text to that simply of expressing “joy”. Recent work attempts to address these issues through a variety of methods, for example by upsampling underrepresented features on which prediction is otherwise sub-optimal (Gesi et al., 2023), or evaluating the importance of input data using shapely values (Karlas et al ˇ ., 2022). However, the mechanistic interpretability work on LLMs to date suggest that our understanding, while improving, is still very limited (e.g. Kramar et al ´ ., 2024; Wu et al., 2023). As such, direct methods for overcoming such biases are, at a minimum, not close at hand. Finally, while much of the focus is naturally on overt racial and gender biases, there may also be pervasive but less observable biases in the content and form of the output. Wendler et al. (2024), for example, provide evidence that that current LLMs trained on large amounts of English text ‘rely on’ English in their latent representations, as if a kind of reference language. One particular area in which the diversity of LLM outputs has been analyzed is on a token-by-token level in the context of decoding strategies. In some situations, using beam search to choose the most likely next token can create degenerate repetitive phrases (Su et al., 2022). Furthermore, a bit like Thelonious Monk’s melodic lines, humans do not string together sequences of the most likely words but occasionally try to surprise the listener by sampling from low-probability words, defying conventions, etc. Holtzman et al. (2020) (referring to Grice, 1975). This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. This paper is available on arxiv under CC BY-NC-SA 4.0 DEED license. available on arxiv available on arxiv

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

Your AI Has a Favorite Opinion—And It’s Not Yours

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Ablative Study on Domain Adapter, Motion Module Design, and MotionLoRA Efficiency

AI for All: Operationalizing Diversity and Inclusion Requirements for AI Systems

Defining Diversity and Inclusion in AI

Concrete Steps for Implementing Diversity and Inclusion in AI

Research Methodology for Operationalizing Diversity and Inclusion Requirements for AI Systems

Operationalizing Diversity and Inclusion Requirements for AI Systems: Results

Ablative Study on Domain Adapter, Motion Module Design, and MotionLoRA Efficiency

AI for All: Operationalizing Diversity and Inclusion Requirements for AI Systems

Defining Diversity and Inclusion in AI

Concrete Steps for Implementing Diversity and Inclusion in AI

Research Methodology for Operationalizing Diversity and Inclusion Requirements for AI Systems

Operationalizing Diversity and Inclusion Requirements for AI Systems: Results

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps