From task structures to world models: what do LLMs know?
Open accessThe DOI matches its Crossref record; 31 of the 31 references it cites were found in a registry or an index.
Abstract
In what sense does a large language model (LLM) have knowledge? We answer by granting LLMs 'instrumental knowledge': knowledge gained by using next-word generation as an instrument. We then ask how instrumental knowledge is related to the ordinary, 'worldly knowledge' exhibited by humans, and explore this question in terms of the degree to which instrumental knowledge can be said to incorporate the structured world models of cognitive science. We discuss ways LLMs could recover degrees of worldly knowledge and suggest that such recovery will be governed by an implicit, resource-rational tradeoff between world models and tasks. Our answer to this question extends beyond the capabilities of a particular AI system and challenges assumptions about the nature of knowledge and intelligence.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Hochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
LoRA: Low-Rank Adaptation of Large Language Models
Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021) · arXiv · doi:10.48550/arxiv.2106.09685 · Cited by 13,453 · cited 1×
“It is of interest to understand how such fi ne-tuning impacts knowledge; however, such procedures typically adapt only certain output stages of the pretrained models, otherwise keeping much of the pretrained weights frozen [33,34].”
BackgroundCognitive maps in rats and men
Tolman, E. C. (1948) · Psychological Review · doi:10.1037/h0061626 · Cited by 6,641
Evolutionary-scale prediction of atomic level protein structure with a language model
Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., et al. (2022) · doi:10.1101/2022.07.20.500902 · Cited by 5,990
Clark, H. H. (1996) · doi:10.1017/cbo9780511620539 · Cited by 5,123
LLaMA: Open and Efficient Foundation Language Models
Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al. (2023) · arXiv · doi:10.48550/arxiv.2302.13971 · Cited by 4,602
Training language models to follow instructions with human feedback
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. (2022) · arXiv · doi:10.48550/arxiv.2203.02155 · Cited by 4,208
Training Language Models to Follow Instructions with Human Feedback
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-2011 · Cited by 3,117
Building machines that learn and think like people
Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017) · Behavioral and Brain Sciences · doi:10.1017/s0140525x16001837 · Cited by 2,550
Building Machines That Learn and Think Like People
Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2016) · arXiv · doi:10.48550/arxiv.1604.00289 · Cited by 2,550
Sparks of Artificial General Intelligence: Early experiments with GPT-4
Bubeck, S., Chandrasekaran, V., Eldan, R., Gehrke, J., Horvitz, E., Kamar, E., Lee, P., Lee, Y. T., Li, Y., Lundberg, S., et al. (2023) · arXiv · doi:10.48550/arxiv.2303.12712 · Cited by 1,526
Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data
Bender, E. M., & Koller, A. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.463 · Cited by 1,082 · cited 1×
“A formal review of this literature is beyond the scope of the present article, but we provide brief pointers in Box 1.”
BackgroundAction understanding as inverse planning
Baker, C. L., Saxe, R., & Tenenbaum, J. B. (2009) · Cognition · doi:10.1016/j.cognition.2009.07.005 · Cited by 920
Lieder, F., & Griffiths, T. L. (2020) · Behavioral and Brain Sciences · doi:10.1017/s0140525x1900061x · Cited by 892
Simulation as an engine of physical scene understanding
Battaglia, P. W., Hamrick, J. B., & Tenenbaum, J. B. (2013) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.1306572110 · Cited by 681
Reworking the language network
Fedorenko, E., & Thompson-Schill, S. L. (2014) · Trends in Cognitive Sciences · doi:10.1016/j.tics.2013.12.006 · Cited by 615 · cited 1×
“Similarly, based on the separation observed in the human brain between language and non-language regions [28,29], researchers have argued [30] that LLMs acquire formal linguistic competence, or knowledge of the rules [30] and statistical regularities [31] of a language, but not ‘ functional…”
BackgroundFunctional specificity for high-level linguistic processing in the human brain
Fedorenko, E., Behr, M. K., & Kanwisher, N. (2011) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.1112937108 · Cited by 595
Introduction to frequency and the emergence of linguistic structure
Bybee, J. L., & Hopper, P. J. (2001) · Typological Studies in Language · doi:10.1075/tsl.45.01byb · Cited by 373 · cited 1×
“…[28,29], researchers have argued [30] that LLMs acquire formal linguistic competence, or knowledge of the rules [30] and statistical regularities [31] of a language, but not ‘ functional linguistic competence ’ , which includes knowledge of and reference to things and processes in the social and…”
BackgroundLearning task-state representations
Niv, Y. (2019) · Nature Neuroscience · doi:10.1038/s41593-019-0470-8 · Cited by 356
Multimodal Few-Shot Learning with Frozen Language Models
Tsimpoukelli, M., Menick, J., Cabi, S., Eslami, S. M. A., Vinyals, O., & Hill, F. (2021) · arXiv · doi:10.48550/arxiv.2106.13884 · Cited by 355
A tutorial introduction to the minimum description length principle
Grunwald, P. (2004) · arXiv · doi:10.48550/arxiv.math/0406077 · Cited by 293
Structuring knowledge with cognitive maps and cognitive graphs
Peer, M., Brunec, I. K., Newcombe, N. S., & Epstein, R. A. (2021) · Trends in Cognitive Sciences · doi:10.1016/j.tics.2020.10.004 · Cited by 283
Eight Things to Know about Large Language Models
Bowman, S. R. (2023) · arXiv · doi:10.48550/arxiv.2304.00612 · Cited by 137
Warren, W. H. (2019) · Journal of Experimental Biology · doi:10.1242/jeb.187971 · Cited by 128 · cited 1×
“The world model in this game comprises the state of the board (for each cell in an 8 × 8 grid, whether it is black, white, or empty) and the set of rules by which each player changes this state.”
BackgroundA theory of learning to infer.
Dasgupta, I., Schulz, E., Tenenbaum, J. B., & Gershman, S. J. (2020) · Psychological Review · doi:10.1037/rev0000178 · Cited by 70
Can Language Models Encode Perceptual Structure Without Grounding? A Case Study in Color
Abdou, M., Kulmizev, A., Hershcovich, D., Frank, S., Pavlick, E., & Søgaard, A. (2021) · arXiv · doi:10.48550/arxiv.2109.06129 · Cited by 70 · cited 1×
“We consider two lines of work: models trained on specialized, non-language domains, and models trained on Internet-scale natural language data.”
MethodCan Language Models Encode Perceptual Structure Without Grounding? A Case Study in Color
Abdou, M., Kulmizev, A., Hershcovich, D., Frank, S., Pavlick, E., & Søgaard, A. (2021) · Proceedings of the 25th Conference on Computational Natural Language Learning · doi:10.18653/v1/2021.conll-1.9 · Cited by 64
Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task
Li, K., Hopkins, A. K., Bau, D., Viégas, F., Pfister, H., & Wattenberg, M. (2022) · arXiv · doi:10.48550/arxiv.2210.13382 · Cited by 61
Wong, L., Grand, G., Lew, A. K., Goodman, N. D., Mansinghka, V. K., Andreas, J., & Tenenbaum, J. B. (2023) · arXiv · doi:10.48550/arxiv.2306.12672 · Cited by 36
Tracr: Compiled Transformers as a Laboratory for Interpretability
Lindner, D., Kramár, J., Farquhar, S., Rahtz, M., McGrath, T., & Mikulik, V. (2023) · arXiv · doi:10.48550/arxiv.2301.05062 · Cited by 8
Cited by
Dissociating language and thought in large language models
Mahowald, K., Ivanova, A. A., Blank, I. A., Kanwisher, N., Tenenbaum, J. B., & Fedorenko, E. (2024) · Trends in Cognitive Sciences · doi:10.1016/j.tics.2024.01.011 · Cited by 345
Zheng, S., Pan, K., Liu, J., & Chen, Y. (2024) · Reliability Engineering & System Safety · doi:10.1016/j.ress.2024.110382 · Cited by 127
Prettyman, A. (2026) · Inquiry · doi:10.1080/0020174x.2024.2439989 · Cited by 18
“For instance, to borrow an example from Yildirim and Paul (2024), when we judge that it is easier to balance a ball on a box than a box on a ball, we use a world model of 3-”
Krauß, V., McGill, M., Kosch, T., Thiel, Y. M., Schön, D., & Gugenheimer, J. (2025) · Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3706598.3713083 · Cited by 13
Di Paolo, L. D., White, B., Guénin-Carlut, A., Constant, A., & Clark, A. (2024) · Philosophical Transactions of the Royal Society B: Biological Sciences · doi:10.1098/rstb.2023.0148 · Cited by 14
“On the one hand, the ‘ground’ of LLMs’ understanding will be explicitly based on the 4E human experience (e.g. [75]).”
Background
Metadata
- Authors
- Yildirim, I., & Paul, L. A.
- Venue
- Trends in Cognitive Sciences
- Publisher
- Elsevier BV
- Published
- 2024-05
- Volume
- 28
- Issue
- 5
- Pages
- 404-415
- ISSN
- 1364-6613 (ISSN-L)
- OpenAlex
- W4392384010
- Cited by
- 87
- References
- 31
- FWCI
- 3.86
- Type
- journal article
- Language
- en
- Open access
- bronze
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.