Dimana Orlinova Anastassova
Roma Tre University
In the evolving landscape of natural language processing (NLP), Large Language Models (LLMs) have achieved unprecedented results, reigniting the debate on whether machines can understand language [12, 6, 3, 2]. However, despite their successes in NLP tasks, scholars believe that LLMs still struggle significantly with natural language understanding (LU) [9, 11].
But what do we mean by language “processing” and “understanding”? Language processing generally involves formal linguistic competence [4, 10], whereas “humanlike” LU often lacks a precise definition [11], with cognitive sciences tending towards operational definitions based on measurable outcomes.
Interestingly, epistemology and philosophy of science offer a broader perspective on understanding as a cognitive achievement linking a subject to an object, examining its nature and relation to knowledge, explanation, and belief. Traditionally, theories of understanding have centered on an idealized subject, which limits their applicability to real-world contexts [1]. Recent philosophical work, however, argues that understanding is not binary but exists in degrees [1, 3], varying across humans’ cognitive developmental stages and potentially in certain animal species. These degrees are often modeled through minimalist [5] and maximalist [7] approaches, yet they have not been systematically integrated with empirical LU evidence.
This leads to our question: Can LLMs achieve some degree of LU, comparable to that attributed to humans? To explore this, we will adopt a minimalist approach: rather than comparing LLMs directly to adults, we will focus on the first degree of LU, found in young children. In fact, by ages 4-5, children are full participants in conversational exchanges, having achieved critical milestones in LU [8, 13, 14], though their understanding remains incomplete [13]. Within this minimalist approach to LU, we focus on how LLMs may achieve a minimal degree of LU based on their formal linguistic competence despite their lack of functional linguistic competence [4, 10].
The present work seeks to bridge empirical evidence with philosophical theories on LU. Its general aim is to move beyond idealized speakers and better account for the existence of degrees of LU, while recognizing that LLMs should not be excluded a priori.
References:
[1] Christoph Baumberger, Claus Beisbart, and Georg Brun. “What Is Understanding? An Overview of Recent Debates in Epistemology and Philosophy of Science”. In: Explaining Understanding: New Perspectives from Epistemology and Philosophy of Science. Ed. by Stephen R. Grimm, Christoph Baumberger, and Sabine Ammon. New York: Routledge, 2016, pp. 1–34. doi: 10.4324/9781315686110.
[2] Pierre Beckmann and Matthieu Queloz. Mechanistic Indicators of Understanding in Large Language Models. 2026. doi: 10 . 48550 / arXiv . 2507 . 08017. arXiv: 2507.08017 [cs].
[3] Yongho Choi. “Linguistic Understanding Beyond Knowledge, Ability, and Process: A Multi-Layered and Multi-Axial Model”. In: Philosophy & Technology 39.50 (2026). doi: 10.1007/s13347-026-01047-y.
[4] Evelina Fedorenko, Anna A. Ivanova, and Tamar I. Regev. “The Language Network as a Natural Kind within the Broader Landscape of the Human Brain”. In: Nature Reviews. Neuroscience 25.5 (2024), pp. 289–312. doi: 10.1038/s41583-024-00802-4.
[5] Stephen R. Grimm. “Understanding and Transparency”. In: Explaining Understanding: New Perspectives from Epistemology and Philosophy of Science. Ed. by Stephen R. Grimm, Christoph Baumberger, and Sabine Ammon. New York: Routledge, 2016, pp. 212–229.
[6] Vladimír Havlík. “Meaning and Understanding in Large Language Models”. In: Synthese 205.9 (2024). doi: 10.1007/s11229-024-04878-4.
[7] Christoph Kelp. “Inquiry, Knowledge and Understanding”. In: Synthese 198.7 (2021), pp. 1583–1593. doi: 10.1007/s11229-018-1803-y.
[8] Patricia K. Kuhl. “Brain Mechanisms in Early Language Acquisition”. In: Neuron 67.5 (2010), pp. 713–727. doi: 10.1016/j.neuron.2010.08.038.
[9] Brenden M. Lake and Gregory L. Murphy. “Word Meaning in Minds and Machines”. In: Psychological Review 130.2 (2023), pp. 401–431. doi: 10.1037/rev0000297.
[10] Kyle Mahowald et al. “Dissociating Language and Thought in Large Language Models”. In: Trends in Cognitive Sciences 28.6 (2024), pp. 517–540. doi: 10.1016/j.tics.2024.01.011.
[11] Melanie Mitchell and David C. Krakauer. “The Debate over Understanding in AI’s Large Language Models”. In: Proceedings of the National Academy of Sciences 120.13 (2023), e2215907120. doi: 10.1073/pnas.2215907120.
[12] Ellie Pavlick. “Symbols and Grounding in Large Language Models”. In: Philosophical Transactions. Series A, Mathematical, Physical, and Engineering Sciences 381.2251: 20220041 (2023). doi: 10.1098/rsta.2022.0041.
[13] Jesse Snedeker. “Children’s Sentence Processing”. In: Sentence Processing. Ed. by Roger P. G. van Gompel. Psychology Press, 2013, pp. 189–220. doi: 10.4324/9780203488454.
[14] George Yule. “First Language Acquisition”. In: The Study of Language. Ed. by George Yule. Cambridge: Cambridge University Press, 2022, pp. 208–227. doi: 10.1017/9781009233446.
Chair: tba
Time:
Location:
SOPhiA 2026 