GPT-ology, Computational Models, Silicon Sampling: How should we think about LLMs in Cognitive Science? (2406.09464v1)

Abstract: LLMs 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, andsilicon sampling" -- and review 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.

• The paper demonstrates GPT-ology as a method to assess LLM cognitive traits through validated psychological tasks.
• It compares LLMs' roles as computational models of human cognition with traditional neural network approaches to reveal strengths and limitations.
• It highlights silicon sampling's potential to simulate human-like outputs while addressing biases and reproducibility challenges.

Exploring Research Paradigms in Cognitive Science with LLMs

Introduction and Background

The paper "GPT-ology, Computational Models, Silicon Sampling: How should we think about LLMs in Cognitive Science?" (2406.09464) provides a comprehensive review of the emerging paradigms in cognitive science research leveraging LLMs. With the increasing prominence of LLMs such as GPT-4 in cognitive science discussions, this paper attempts to categorize the variable approaches and insights gained from these models in understanding both artificial and human cognition.

Research Paradigms

GPT-ology

GPT-ology involves analyzing the cognitive capabilities and limitations of specific LLMs, like GPT variants. It draws inferences about the state and evolution of artificial cognition. Researchers often paper LLMs' responses to validated psychological scales or tasks, inferring "traits" like personality or biases that might influence generated outputs. These studies also highlight certain inaccuracies or human-like errors LLMs perform.

Figure 1: Various inferences drawn from identical experimental setups assessing LLMs' capabilities, illustrating the diverse research paradigms.

Critically, the anthropomorphic interpretation of traits in LLMs and the reliability concerning model updates pose significant challenges to the scientific inference within this paradigm. The brittleness and susceptibility of LLMs to prompt variations are particularly noteworthy, suggesting potential data memorization rather than robust cognitive processing.

LLMs as Computational Models

LLMs serve as computational analogs to human cognition, akin to prior comparisons with simpler neural network models. Their scale introduces opportunities for deeper experimentation, potentially revealing insights into learning mechanisms, biases, and moral reasoning when juxtaposed with human patterns. Researchers leverage LLMs for "existence proofs" about the sufficiency of conditions for cognitive capabilities, focusing on learned event knowledge or symbolic reasoning from statistical language inputs.

However, inferring human cognition models from LLM outputs brings complexities, particularly when accounting for training data presence in LLMs or drawing conclusions from failures (null results). The paper cautions against asymmetrical existence proof logic as cognitive insights can be misleading if model limitations are misconstrued.

Silicon Sampling

Silicon sampling utilizes LLMs to simulate human behaviors and responses, exploring fidelity in replicating human-like outputs. Despite offering simulations of hard-to-recruit populations, biases in training data and fidelity in approximation of human response distributions present concerns. Misinterpretations could arise if LLMs, perceived as mimicking individual humans, are instead aggregating population-level knowledge.

Challenges and Outstanding Issues

The paper delineates critical issues facing LLM research:

1. Model Selection: With numerous LLM variants, the choice of model impacts reproducibility and generalizability.
2. Inference Challenges: Differentiating cognitive insights from engineering findings when varying LLM capabilities is difficult.
3. Proprietary Nature: Closed-source models constrain transparency and consistency in scientific inference.
4. Training Data: Lack of visibility and representation biases in LLM training data contribute to unreliable scientific claims.
5. Reproducibility: Stochasticity and prompt engineering complexities demand new scientific conventions for reliable experimentation.
6. Active Development: The rapid evolution of LLMs potentially renders research findings transient.

Conclusion and Future Directions

The exploration of LLMs within cognitive science has introduced novel paradigms for understanding cognition, yet fundamental challenges remain in ensuring robust, stable scientific inference. Future research endeavors should prioritize models with transparent methodologies and address the outstanding scientific and ethical issues highlighted. The paper serves as a catalyst for ongoing discourse aimed at refining and evolving cognitive science methodologies in the LLM domain.