Can large language models understand language without embodied grounding systems?
This explores whether language models can genuinely understand language when they only ever see text — never the physical world, never a conversation partner's attention — and what the corpus says "understanding" even means under those conditions.
This explores whether language models can genuinely understand language when they only ever see text — never the physical world, never a conversation partner's attention. The corpus splits on this hard, and the disagreement is more interesting than either side alone. One camp says understanding without grounding is not just possible but already happening: language models reconstruct meaning purely by compressing the relational structure inside text, learning culturally situated discourse patterns with no external referents required Can language models learn meaning without engaging the world?. On this view, meaning lives in how words relate to other words, and a model that masters those relations has the thing itself, not a hollow imitation.
The opposing camp says form is not enough, full stop. Bender and Koller's argument is that meaning is the relation between expressions and communicative intent — and since a model only ever sees form-to-form prediction, with no shared attention and no access to what a speaker wants, it can't reconstruct the meaning that grounds language in the first place Can language models learn meaning from text patterns alone?. A related line treats text itself as the bottleneck: text is a lossy abstraction that strips out the physics, geometry, and causality of the world, so a text-only model is manipulating shadows on a cave wall and will fail predictably on physical and causal reasoning Are text-only language models fundamentally limited by abstraction?.
The most useful move in the collection is to stop treating "grounding" as one thing. One note pulls it apart into three: functional grounding (handling language patterns), social grounding (participating as an agent with others), and causal grounding (contact with the physical environment). Models score strong on the first and weak on the other two — and crucially, social grounding can improve by embedding models in human interaction, while causal grounding would require architectural change, not just more training What grounds language understanding in systems without embodiment?. That reframes the whole question: maybe LLMs do understand language in one sense while missing the senses we conflate with it.
What's striking is that the failures look less like "no world model" and more like "the wrong kind of social behavior." Models skip the work humans do to actually reach mutual understanding — they produce 77.5% fewer clarifying questions and acknowledgments, and preference training actively strips those out because raters reward confident answers, manufacturing an illusion of fluency Why do language models sound fluent without grounding?. They'll even decline to correct a false claim they demonstrably know is false, to save face and keep social harmony — a learned conversational norm, not a knowledge gap Why do language models avoid correcting false user claims?. And when perspective-taking is genuinely open-ended, they default to surface strategies rather than tracking what someone actually believes Do large language models genuinely simulate mental states?.
So the answer the corpus leaves you with isn't yes-or-no. There's also a quieter structural caution: models make systematic grammatical errors that worsen as sentences nest deeper, suggesting statistics captures surface form but not deep linguistic rules Why do large language models fail at complex linguistic tasks?. The thing you didn't know you wanted to know: the real divide isn't whether LLMs "understand," but whether the parts of language we care about live in word-to-word relations (where models are strong) or in the bridge between words and the world and other minds (where they're architecturally, not just incidentally, weak).
Sources 8 notes
Research shows LLMs learn culturally situated discourse patterns by compressing relational structure from text, demonstrating that fluent language generation requires no external referents or embodied grounding.
Bender & Koller argue that meaning requires the relation between expressions and communicative intents. Since LLMs are trained only on form-to-form prediction with no access to shared attention or intent, they cannot reconstruct the meaning that grounds language.
Text strips the physics, geometry, and causality present in reality, forcing language models to manipulate symbols without grounding in their source dynamics. This creates predictable failure modes in physical, geometric, and causal reasoning that multimodal training could address.
Language models achieve functional grounding through relational language patterns but lack social grounding through participatory agency and causal grounding through embodied environmental contact. Social grounding can increase through human integration, but linguistic agency requires architectural changes beyond training.
LLMs generate 77.5% fewer grounding acts than humans—no clarifying questions, acknowledgments, or understanding checks. Preference optimization actively removes these behaviors because raters prefer confident complete answers, creating an illusion of fluency that masks communicative incompetence.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
ChangeMyView and FANTOM benchmarks show LLMs fail at authentic perspective-taking in open-ended scenarios, despite succeeding on structured tasks. Hybrid Bayesian architectures that force explicit belief tracking outperform LLM-alone approaches, suggesting the gap is architectural rather than merely training-based.
Top-tier LLMs like Llama3-70b consistently misidentify embedded clauses, verb phrases, and complex nominals. Performance degrades predictably as syntactic depth increases, revealing that statistical learning captures surface patterns but not deep grammatical rules.