Cole Robertson

Research

I use the experimental toolkit of cognitive science to study how minds, human and artificial, represent the world and act on it.

World models in large language models

Do language models reason over internal models of the world, or over statistical associations? With Philip Wolff at Emory, I adapted mental-models paradigms from cognitive science to test this, using 810 novel pulley-system problems kept off the public web so that no model could have seen them. Four frontier models tracked the causally relevant variables but estimated mechanical advantage by counting pulleys, and fell to chance when success required representing how a system is connected (arXiv:2507.15521). Current work uses functional localization and targeted ablation to find the representations behind this behavior in open-weight models.

Multi-agent safety

With Soumya Banerjee at the University of York, I am developing work on emergent goals in populations of AI agents: whether agents form unspecified subgoals when they transfer solutions to new environments, and whether populations of narrowly tasked agents drift toward a shared objective nobody set. This work is the subject of a proposal now under review with Schmidt Sciences. A related project asks whether the conditions that produced prosocial behavior in human evolution can be recreated in the training environments of artificial agents.

Language, cognition, and decision-making

My doctoral work, supervised by Robin Dunbar, Asifa Majid, and Seán Roberts, asked whether the way a language talks about the future shapes how its speakers value the future. Experiments in English and Dutch showed that low-certainty modal verbs, not future tenses, increase psychological discounting (Cognitive Science, 2023; Cognition, 2026). The full thesis is available here.

Explainable AI for clinical speech

A fine-tuned transformer predicted amyloid status in primary progressive aphasia from short samples of connected speech with 92% accuracy, ten points above prior work, and explanation models surfaced candidate linguistic markers of the pathology (npj Dementia, 2026). At Ogma Therapy I built phonetic speech recognition for children with speech and language disorders.

Publications

In preparation

  1. Robertson, C., & Wolff, P. Functional localization of emergent world-model reasoning in frontier large language models.
  2. Robertson, C., & Dunbar, R. I. M. Linguistic recursion and the depth of embedded theory of mind in humans and large language models.
  3. Robertson, C. Ecologies of human–machine interdependence: Selecting for cross-boundary mutualism in multi-agent systems.
  4. Robertson, C., Tomanek, K., & Hain, T. A novel transfer-learning approach for connectionist temporal classification to estimate goodness of pronunciation in language-impaired children.

Published

  1. 2026Robertson, C., Hochberg, D., Quimby, M., Wolff, P., Dickerson, B. C., & Rezaii, N. (2026). Predicting amyloid status in primary progressive aphasia using explainable artificial intelligence. npj Dementia, 2, Article 41. https://doi.org/10.1038/s44400-026-00092-w
  2. 2026Robertson, C., Roberts, S. G., Majid, A., Lu, T., Wolff, P., & Dunbar, R. I. M. (2026). Low-certainty modals not future tenses cause increased psychological discounting in English relative to Dutch. Cognition, 267, Article 106338. https://doi.org/10.1016/j.cognition.2025.106338
  3. 2025Robertson, C., & Wolff, P. (2025). LLM world models are mental: Output layer evidence of brittle world model use in LLM mechanical reasoning. In Proceedings of the Superintelligence Conference (SiC). Full paper: arXiv:2507.15521
  4. 2025Robertson, C., Roberts, S. G., Majid, A., & Dunbar, R. I. M. (2025). Language and economic behaviour: Future tense use causes less not more temporal discounting. PLOS ONE, 20(5), e0317422. https://doi.org/10.1371/journal.pone.0317422
  5. 2023Robertson, C., & Roberts, S. G. (2023). Not when but whether: Modality and future time reference in English and Dutch. Cognitive Science, 47(1), e13224. https://doi.org/10.1111/cogs.13224
  6. 2023Robertson, C., Carney, J., & Trudell, S. (2023). Language about the future on social media as a novel marker of anxiety and depression: A big-data and experimental analysis. Current Research in Behavioral Sciences, 4, 100104. https://doi.org/10.1016/j.crbeha.2023.100104
  7. 2022Carney, J., & Robertson, C. (2022). Five studies evaluating the impact on mental health and mood of recalling, reading, and discussing fiction. PLOS ONE, 17(4), e0266323. https://doi.org/10.1371/journal.pone.0266323
  8. 2021Gotti, G., Roberts, S. G., Fasan, M., & Robertson, C. (2021). Language in economics and accounting research: The role of linguistic history. The International Journal of Accounting, 56(3), 2150015. https://doi.org/10.1142/S1094406021500153
  9. 2021Robertson, C. (2021). The future is a thing of possibility: Cross-linguistic differences in modal future time reference affect (risky) intertemporal decisions. Oxford: Holywell Press Limited. Thesis page and PDF
  10. 2019Carney, J., Robertson, C., & Dávid-Barrett, T. (2019). Fictional narrative as a variational Bayesian method for estimating social dispositions in large groups. Journal of Mathematical Psychology, 93, 102279. https://doi.org/10.1016/j.jmp.2019.102279
  11. 2018Carney, J., & Robertson, C. (2018). People searching for meaning in their lives find literature more engaging. Review of General Psychology, 22(2), 199–209. https://doi.org/10.1037/gpr0000134
  12. 2018Dunbar, R. I. M., MacCarron, P., & Robertson, C. (2018). Trade-off between fertility and predation risk drives a geometric sequence in the pattern of group sizes in baboons. Biology Letters, 14(3), 20170700. https://doi.org/10.1098/rsbl.2017.0700
  13. 2017Robertson, C., Tarr, B., Kempnich, M., & Dunbar, R. (2017). Rapid partner switching may facilitate increased broadcast group size in dance compared with conversation groups. Ethology, 123(10), 736–747. https://doi.org/10.1111/eth.12642
  14. 2017Dunbar, R. I. M., Launay, J., Wlodarski, R., Robertson, C., Pearce, E., Carney, J., … (2017). Functional benefits of (modest) alcohol consumption. Adaptive Human Behavior and Physiology, 3(2), 118–133. https://doi.org/10.1007/s40750-016-0058-4
  15. 2017Dunbar, R., Launay, J., Pearce, E., Wlodarski, R., Carney, J., MacCarron, P., Robertson, C. (2017). Friends on tap: The role of pubs at the heart of the community. St. Albans, UK: Campaign for Real Ale.

Talks and presentations

  1. 2025Robertson, C., & Wolff, P. (2025). Cognitive-science approaches to evaluating mechanical reasoning in LLMs: Output layer evidence of shallow world-model use [Accepted talk]. Superintelligence Conference (SiC).
  2. 2025Robertson, C., & Wolff, P. (2025). LLMs have “mental” models: Latent world models in LLM network weights can be inferred from output layer inference [Accepted poster]. Cognitive Science.
  3. 2024Robertson, C. (2024). How would industry like to engage with universities to develop AI [Panel discussion]. Centre for Human-Inspired AI, University of Cambridge.
  4. 2023Robertson, C. (2023). Using cognitive science methods to probe latent representations in large language models. Working paper series, Emory University.
  5. 2020Robertson, C. (2020). Not when but if: Probability not time determines cross-linguistically driven differences in psychological discounting. Cognitive Linguistics.