Potemkin Understanding in Large Language Models (2025)

A paper posted to arXiv on June 26, 2025 introduces a formal framework for judging whether large language model benchmark performance reflects genuine understanding, arguing that benchmarks such as AP exams are valid tests only if LLMs misunderstand concepts the way humans do. The authors, including Marina Mancoridis, present two procedures for quantifying these "potemkin" understandings — one using a purpose-built benchmark across three domains and one general procedure giving a lower bound on their prevalence — and report that potemkins are ubiquitous across models, tasks, and domains. The paper states these failures reflect deeper internal incoherence in concept representations, not merely incorrect understanding.

Computer Science Computation and Language Submitted on 26 Jun 2025 v1 https://arxiv.org/abs/2506.21521v1 , last revised 29 Jun 2025 this version, v2 Title:Potemkin Understanding in Large Language Models View PDF https://arxiv.org/pdf/2506.21521 HTML experimental https://arxiv.org/html/2506.21521v2 Abstract:Large language models LLMs are regularly evaluated using benchmark datasets. But what justifies making inferences about an LLM's capabilities based on its answers to a curated set of questions? This paper first introduces a formal framework to address this question. The key is to note that the benchmarks used to test LLMs -- such as AP exams -- are also those used to test people. However, this raises an implication: these benchmarks are only valid tests if LLMs misunderstand concepts in ways that mirror human misunderstandings. Otherwise, success on benchmarks only demonstrates potemkin understanding: the illusion of understanding driven by answers irreconcilable with how any human would interpret a concept. We present two procedures for quantifying the existence of potemkins: one using a specially designed benchmark in three domains, the other using a general procedure that provides a lower-bound on their prevalence. We find that potemkins are ubiquitous across models, tasks, and domains. We also find that these failures reflect not just incorrect understanding, but deeper internal incoherence in concept representations. Submission history From: Marina Mancoridis view email https://arxiv.org/show-email/15d8ef36/2506.21521 Thu, 26 Jun 2025 17:41:35 UTC 2,403 KB \ v1\ https://arxiv.org/abs/2506.21521v1 v2 Sun, 29 Jun 2025 18:12:45 UTC 2,402 KB References & Citations Loading... Bibliographic and Citation Tools Bibliographic Explorer What is the Explorer? https://info.arxiv.org/labs/showcase.html arxiv-bibliographic-explorer Connected Papers What is Connected Papers? https://www.connectedpapers.com/about Litmaps What is Litmaps? https://www.litmaps.co/ scite Smart Citations What are Smart Citations? https://www.scite.ai/ Code, Data and Media Associated with this Article alphaXiv What is alphaXiv? https://alphaxiv.org/ CatalyzeX Code Finder for Papers What is CatalyzeX? https://www.catalyzex.com DagsHub What is DagsHub? https://dagshub.com/ Gotit.pub What is GotitPub? http://gotit.pub/faq Hugging Face What is Huggingface? https://huggingface.co/huggingface ScienceCast What is ScienceCast? https://sciencecast.org/welcome Demos Recommenders and Search Tools Influence Flower What are Influence Flowers? https://influencemap.cmlab.dev/ CORE Recommender What is CORE? https://core.ac.uk/services/recommender arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs https://info.arxiv.org/labs/index.html .