Do Large Language Models know what humans know?

Abstract: Humans can attribute beliefs to others. However, it is unknown to what extent this ability results from an innate biological endowment or from experience accrued through child development, particularly exposure to language describing others' mental states. We test the viability of the language exposure hypothesis by assessing whether models exposed to large quantities of human language display sensitivity to the implied knowledge states of characters in written passages. In pre-registered analyses, we present a linguistic version of the False Belief Task to both human participants and a LLM, GPT-3. Both are sensitive to others' beliefs, but while the LLM significantly exceeds chance behavior, it does not perform as well as the humans, nor does it explain the full extent of their behavior -- despite being exposed to more language than a human would in a lifetime. This suggests that while statistical learning from language exposure may in part explain how humans develop the ability to reason about the mental states of others, other mechanisms are also responsible.

• The paper demonstrates that GPT-3 attains 74.5% accuracy in belief attribution tasks, though it falls short of human performance at 82.7%.
• The study employs a modified False Belief Task to compare belief sensitivity between GPT-3 and human participants.
• The findings imply that while language exposure enhances belief attribution, LLMs may require additional cognitive mechanisms to fully mirror human social cognition.

An Analysis of the Role of Language Exposure in Belief Attribution: A Study of LLMs

The paper "Do LLMs know what humans know?" by Sean Trott et al. explores the extent to which LLMs, specifically GPT-3, exhibit sensitivity to the belief states of characters in text, akin to human belief attribution. The study centers on the exploration of whether exposure to language can independently drive the development of belief sensitivity, a key component of human social cognition, without the need for innate or experiential mechanisms traditionally thought necessary.

Methodological Approach

The authors employed a linguistic adaptation of the False Belief Task (FBT), a benchmark for evaluating belief attribution capabilities, to assess both GPT-3 and human participants. This task involves narrative scenarios where an object is moved without a character's knowledge, and participants predict where the character will search for the object. The manipulation of the character’s knowledge state—whether 'True Belief' (aware of the object's new location) or 'False Belief' (unaware)—forms the crux of the task. Crucially, both explicit and implicit prompts were used to test the belief sensitivity of participants and GPT-3.

Key Findings

GPT-3 demonstrated significant sensitivity to the task's belief manipulations, achieving an accuracy of 74.5% using log-odds predictions and 73.4% in generative predictive tasks. However, this performance was lower than that of human participants, who correctly attributed belief states 82.7% of the time. Importantly, the statistical model of human behavior indicated that GPT-3's outputs did not fully account for human performance, suggesting additional cognitive mechanisms at play in humans that LLMs do not capture solely through language exposure.

Implications

Theoretical Insights:

1. Distributional Learning: The study posits a partial support for the hypothesis that linguistic structure can capture some aspects of belief attribution. Yet, the gap between GPT-3 and human performance underscores the potential necessity of innate cognitive frameworks or non-linguistic experiences.
2. Model Comparison: The experiment with varying model sizes and training data suggests that even larger models might close the performance gap. This raises questions on scalability and the marginal returns of increasing model parameters relative to training data amount.

Practical Considerations:

1. AI Design: Insights from LLMs can inform the design of more sophisticated AI systems with enhanced social interaction capabilities, albeit recognizing the current limitations in achieving full human-equivalent theory of mind faculties.
2. Task Validity: The ability of LLMs to perform the FBT raises questions about the validity of this task as a standalone measure of theory of mind, potentially prompting revisions in how such cognitive capacities are tested.

Future Directions

Given the demonstrated limitations of LLMs like GPT-3 in achieving comprehensive belief attribution, further research could:

• Investigate whether scaling models in terms of both parameters and diverse linguistic contexts might improve performance in tasks related to cognitive phenomena like belief attribution.
• Explore cross-disciplinary approaches, integrating cognitive science and machine learning to unveil novel architectures that could better simulate human social cognition.
• Develop empirical methodologies to test the influence of multimodal inputs and embodied experiences on belief attribution, potentially using hybrid models.

In summary, while LLMs show potential in approximating certain components of human cognitive abilities from language alone, the observed discrepancies highlight the complex interplay of linguistic input with underlying cognitive faculties. As AI continues to evolve, the quest to bridge these gaps offers promising avenues for enhanced models that are more akin to human social cognition.