Does semantic grounding in language models come in degrees?
Rather than asking whether LLMs truly understand meaning, this explores whether grounding is actually a multi-dimensional spectrum. The question matters because it reframes the sterile understand/don't-understand debate into measurable, distinct capacities.
The grounding debate — "do LLMs understand meaning?" — has been stuck in a binary framing that obscures more than it reveals. "Understanding AI" (Schneider 2024) proposes that semantic grounding is multi-criterial and gradual, distinguishing three fundamentally different types:
Functional grounding (system-internal): The capacity to integrate linguistic expressions into a coherent functional network of world knowledge and inference. LLMs have this strongly — their internal representations are structurally coherent, they form implicit world models, and their outputs show sophisticated semantic integration. This is purely about what happens inside the system boundary.
Social grounding (community-embedded): The capacity to participate in the "language games" of a linguistic community (Wittgenstein). Social grounding is constituted by the functional roles an agent plays through behavior in the world — "long-arm" functional roles. LLMs have this weakly. The more they are used as communicative partners and integrated into linguistic practices, the more they acquire it — Can LLMs acquire social grounding through linguistic integration?.
Causal grounding (world-connected): Direct contact with the physical world that language refers to. LLMs lack this directly. However, training data is produced by causally grounded humans, opening the door to Can large language models develop genuine world models without direct environmental contact? — an indirect causal grounding functionally established through world model formation.
The key insight: grounding is not a yes/no property but a graduated, multi-dimensional one. Arguments that "LLMs don't understand" often implicitly invoke causal grounding criteria; arguments that "LLMs do understand" typically invoke functional grounding criteria. Both are grounding, but different kinds.
This taxonomy connects to Do language models actually use their encoded knowledge? — which is a functional-to-causal gap: something can be encoded (functional) without influencing what is generated (and without causal world connection).
Peirce's semiotic framework provides a complementary vocabulary that maps precisely onto this tri-partite structure. Functional grounding corresponds to symbolic Thirdness (generalization, abstraction, pattern inference — the internal symbol system). Causal grounding corresponds to Secondness (direct encounter with brute fact, world resistance). Social grounding corresponds to socially-mediated Thirdness (shared, negotiated meaning within a community of interpreters). The Can AI systems achieve real alignment without world contact? insight follows from this mapping: a system operating in functional/Thirdness alone without Secondness and socially-mediated Thirdness cannot guarantee its goal representations track real-world values. Tool-use and RAG introduce what that paper calls "proto-indexicality" — beginning to address the Secondness/causal gap without fully closing it.
Source: Linguistics, NLP, NLU
Related concepts in this collection
-
Can LLMs acquire social grounding through linguistic integration?
Explores whether LLMs gradually develop social grounding as they become embedded in human language practices, analogous to child language acquisition. Tests whether grounding is a fixed property or an outcome of participatory use.
the social grounding dimension
-
Can large language models develop genuine world models without direct environmental contact?
Do LLMs extract meaningful world structures from human-generated text despite lacking direct sensory access to reality? This matters for understanding what kind of grounding and knowledge these systems actually possess.
the causal grounding dimension
-
Do language models actually use their encoded knowledge?
Probes can detect that LMs encode facts internally, but do those encoded facts causally influence what the model generates? This explores the gap between knowing and doing.
functional-causal gap
-
Do humans and LLMs differ fundamentally or just superficially?
Explores whether the gap between human and AI cognition is categorical or contextual. Matters because it shapes how we design, evaluate, and interact with language models in practice.
participant perspective tracks social/functional grounding; observer perspective tracks causal grounding
-
Do large language models reason symbolically or semantically?
Can LLMs follow explicit logical rules when those rules contradict their training knowledge? Testing whether reasoning operates independently of semantic associations reveals what computational mechanisms actually drive LLM multi-step inference.
functional grounding through semantic associations explains why LLMs reason successfully within commonsense boundaries but fail when semantics are decoupled from logic
-
Do vector embeddings actually measure task relevance?
Vector embeddings rank semantic similarity, but RAG systems need topical relevance. When these diverge—as with king/queen versus king/ruler—does similarity-based retrieval fail in production?
the retrieval instantiation of the functional-causal gap: vector embeddings capture functional grounding (co-occurrence associations) but not task relevance (what actually answers the query); the king/queen/ruler problem is exactly functional association without causal connection to query intent
-
Can AI systems learn social norms without embodied experience?
Large language models exceed individual human accuracy at predicting collective social appropriateness judgments. Does this reveal that embodied experience is unnecessary for cultural competence, or do systematic AI failures point to limits of statistical learning?
forces a revision of "social grounding is weak": LLMs predict social norms at the 100th percentile, suggesting functional grounding may be sufficient for social-norm competence even without social participation; the "weak" rating may apply to social participation but not to social prediction
-
Do foundation models learn world models or task-specific shortcuts?
When transformer models predict sequences accurately, are they building genuine world models that capture underlying physics and logic? Or are they exploiting narrow patterns that fail under distribution shift?
task-specific heuristics are the behavioral signature of functional grounding operating without causal grounding: models learn to predict legal next states within training distributions (strong functional integration) but cannot generalize via underlying world structure (absent causal grounding), forcing reliance on narrow non-transferable procedures
Click a node to walk · click center to open · click Open full network for a force-directed map
Original note title
llm semantic grounding is tri-partite functional grounding is strong social grounding is weak causal grounding is indirect