arXiv:2406.09464v1 Announce Type: new
Abstract: Large Language Models have taken the cognitive science world by storm. It is perhaps timely now to take stock of the various research paradigms that have been used to make scientific inferences about “cognition” in these models or about human cognition. We review several emerging research paradigms — GPT-ology, LLMs-as-computational-models, and “silicon sampling” — and review recent papers that have used LLMs under these paradigms. In doing so, we discuss their claims as well as challenges to scientific inference under these various paradigms. We highlight several outstanding issues about LLMs that have to be addressed to push our science forward: closed-source vs open-sourced models; (the lack of visibility of) training data; and reproducibility in LLM research, including forming conventions on new task “hyperparameters” like instructions and prompts.
Understanding Large Language Models: A Multidisciplinary Analysis
Large Language Models (LLMs) have revolutionized the field of cognitive science, prompting researchers to examine their potential implications for both artificial and human cognition. In this article, we aim to explore the multi-disciplinary nature of the concepts surrounding LLM research. We will analyze three key research paradigms that have emerged in the study of LLMs: GPT-ology, LLMs-as-computational-models, and “silicon sampling”.
GPT-ology: Unraveling the workings of Large Language Models
The GPT-ology paradigm focuses on understanding the internal mechanisms and capabilities of LLMs, such as GPT (Generative Pre-trained Transformer). Researchers employing this paradigm aim to uncover the underlying cognitive processes and representations encoded within LLMs. By examining the behavior and performance of these models on various tasks, they strive to draw insights about their cognitive abilities.
One challenge faced in GPT-ology is the lack of transparency in closed-source models. Transparency is crucial for better understanding LLMs and verifying claims made about their cognitive capabilities. The research community must advocate for increased access to the inner workings of these models and the training data they rely on.
LLMs-as-computational-models: Bridging the gap between artificial and human cognition
The LLMs-as-computational-models paradigm aims to use LLMs as tools to study and simulate human cognitive processes. Researchers employing this paradigm explore the similarities and differences between LLM performance and human cognitive abilities. By leveraging the computational power of LLMs, cognitive scientists can investigate complex cognitive phenomena with greater speed and scale.
One critical issue raised in this paradigm is how to ensure the reliability and reproducibility of LLM research. Reproducibility is crucial for establishing the validity of findings and building upon existing knowledge. The scientific community needs to establish conventions for new task “hyperparameters,” such as instructions and prompts, to ensure consistency in experiments and allow for meaningful comparisons across studies.
“Silicon sampling”: Utilizing LLMs to generate novel insights
“Silicon sampling” refers to the practice of using LLMs to generate synthetic data or simulate cognitive phenomena. Researchers employing this approach leverage LLMs’ generation capabilities to explore novel hypotheses, design experiments, and examine phenomena that are challenging to observe directly. By generating new data and simulations, they can test and refine theories in a controlled environment.
A critical consideration in “silicon sampling” is the ethical use of LLMs. These models have the potential to create highly realistic text and media, raising concerns about misinformation, bias, and malicious uses. Guidelines and safeguards must be established to ensure responsible and ethical use of LLMs in generating synthetic data or simulations.
Future Directions and Outstanding Issues
As the field of LLM research progresses, several outstanding issues must be addressed for further advancements. Firstly, increasing the transparency of LLMs, particularly through open-sourced models, will foster better scrutiny and understanding of their capabilities. Secondly, the availability and visibility of training data are crucial for replicating and building upon LLM research. Efforts should be made to make training data more accessible while respecting privacy concerns and data ownership rights. Lastly, establishing conventions for task hyperparameters in LLM research, such as instructions and prompts, will enhance comparability across studies and ensure robust scientific inference.
By recognizing the multidisciplinary nature of LLM research paradigms and addressing the outstanding issues, we can propel the field forward and unlock new insights into both artificial and human cognition. Collaborations between cognitive scientists, computer scientists, ethicists, and other relevant disciplines will play a vital role in advancing this fascinating area of research.