Language Statistics and False Belief Reasoning: Evidence from 41 Open-Weight LMs
arXiv:2602.16085v1 Announce Type: new Abstract: Research on mental state reasoning in language models (LMs) has the potential to inform theories of human social cognition--such as the theory that mental state reasoning emerges in part from language exposure--and our understanding of LMs themselves. Yet much published work on LMs relies on a relatively small sample of closed-source LMs, limiting our ability to rigorously test psychological theories and evaluate LM capacities. Here, we replicate and extend published work on the false belief task by assessing LM mental state reasoning behavior across 41 open-weight models (from distinct model families). We find sensitivity to implied knowledge states in 34% of the LMs tested; however, consistent with prior work, none fully ``explain away'' the effect in humans. Larger LMs show increased sensitivity and also exhibit higher psychometric predictive power. Finally, we use LM behavior to generate and test a novel hypothesis about human cognit
arXiv:2602.16085v1 Announce Type: new Abstract: Research on mental state reasoning in language models (LMs) has the potential to inform theories of human social cognition--such as the theory that mental state reasoning emerges in part from language exposure--and our understanding of LMs themselves. Yet much published work on LMs relies on a relatively small sample of closed-source LMs, limiting our ability to rigorously test psychological theories and evaluate LM capacities. Here, we replicate and extend published work on the false belief task by assessing LM mental state reasoning behavior across 41 open-weight models (from distinct model families). We find sensitivity to implied knowledge states in 34% of the LMs tested; however, consistent with prior work, none fully ``explain away'' the effect in humans. Larger LMs show increased sensitivity and also exhibit higher psychometric predictive power. Finally, we use LM behavior to generate and test a novel hypothesis about human cognition: both humans and LMs show a bias towards attributing false beliefs when knowledge states are cued using a non-factive verb (``John thinks...'') than when cued indirectly (``John looks in the...''). Unlike the primary effect of knowledge states, where human sensitivity exceeds that of LMs, the magnitude of the human knowledge cue effect falls squarely within the distribution of LM effect sizes-suggesting that distributional statistics of language can in principle account for the latter but not the former in humans. These results demonstrate the value of using larger samples of open-weight LMs to test theories of human cognition and evaluate LM capacities.
Executive Summary
This article presents an extensive study on the mental state reasoning abilities of 41 open-weight language models (LMs) in relation to false belief reasoning. By leveraging a larger sample of open-source LMs, the study demonstrates the value of this approach in testing theories of human cognition and evaluating LM capacities. The findings reveal that LMs exhibit sensitivity to implied knowledge states, albeit to a lesser extent than humans. Furthermore, the study uncovers a novel hypothesis about human cognition, suggesting that humans and LMs both exhibit a bias towards attributing false beliefs when knowledge states are cued using a non-factive verb. The results have significant implications for our understanding of the relationship between language and cognition, as well as the design and evaluation of LMs.
Key Points
- ▸ The study assesses the mental state reasoning abilities of 41 open-weight LMs in relation to false belief reasoning.
- ▸ Larger LMs show increased sensitivity and exhibit higher psychometric predictive power.
- ▸ The study uncovers a novel hypothesis about human cognition, suggesting a bias towards attributing false beliefs when knowledge states are cued using a non-factive verb.
Merits
Strength
The study utilizes a larger sample of open-source LMs, providing a more comprehensive understanding of LM capacities and human cognition.
Demerits
Limitation
The study relies on a limited number of open-source LMs, which may not be representative of the broader LM landscape.
Expert Commentary
The study presents a significant contribution to the field of natural language processing, offering a more comprehensive understanding of the mental state reasoning abilities of LMs. The use of a larger sample of open-source LMs provides a more robust evaluation of LM capacities and informs LM design. However, the study's reliance on a limited number of open-source LMs may limit its generalizability. Future studies should aim to replicate these findings with a more diverse range of LMs. Additionally, the study's implications for human cognition and LM design warrant further exploration.
Recommendations
- ✓ Future studies should aim to replicate these findings with a more diverse range of LMs.
- ✓ The study's implications for human cognition and LM design warrant further exploration and development of novel applications.
