How does syntactic encoding relate to semantic feature representation?
This explores whether the way LLMs encode grammar (sentence structure, syntactic relations) is separate from, or tangled up with, the way they encode meaning (semantic features) — and what the corpus says about how these two layers sit alongside each other inside a model.
This question is really asking: when a language model handles structure (grammar, word order, clause nesting) versus meaning (what words actually refer to or feel like), are those two different systems, or one blurred-together thing? The corpus suggests they're encoded by genuinely distinct geometric machinery — but with meaning often quietly overriding structure when the two compete.
On the syntax side, models build surprisingly clean structured geometry. The Polar Probe work finds that LLMs encode syntactic relations in polar coordinates — distance captures *how strongly* two words relate, and angle captures the *type and direction* of the relation, nearly doubling accuracy over distance-only methods How do language models encode syntactic relations geometrically?. Models can even step outside behavior into genuine analysis: with chain-of-thought, they construct valid syntactic trees and phonological rules, not just produce grammatical sentences Can language models actually analyze language structure?. So syntactic structure is really *in there*, represented in a near-symbolic way.
Meaning lives in its own geometry, and a different shape. Semantic features in embeddings collapse onto just three principal axes that mirror the human 'evaluation-potency-activity' structure — and because they're entangled, nudging one feature predictably drags aligned ones along Do LLM semantic features organize along human evaluation dimensions?. Even static embeddings, before attention does any work, already carry rich lexical meaning — valence, concreteness, iconicity Do transformer static embeddings actually encode semantic meaning?. And circuit tracing shows the whole stack is layered: token-level inputs at the bottom, abstract concepts in the middle, functional operations near the top How do language models organize features across processing layers?. Syntactic detail tends to be the low/middle-tier scaffolding; semantic abstraction is what the upper tiers reach for.
The revealing part is what happens when structure and meaning pull against each other — meaning usually wins, often to the model's detriment. LLMs systematically prefer high-frequency surface phrasings over semantically identical rare paraphrases, tracking statistical mass rather than recognizing meaning Do language models really understand meaning or just surface frequency?. When you strip semantic content out of a reasoning task and leave only the formal rules, performance collapses — the model was leaning on token associations, not syntactic logic Do large language models reason symbolically or semantically?. And their grasp of structure frays exactly where it should hold: as syntactic depth increases — embedded clauses, complex nominals — top models reliably misidentify the grammar Why do large language models fail at complex linguistic tasks?.
The thing you might not have known you wanted to know: a model can hold a clean, almost mathematical map of syntax and *still* fall back on semantic/statistical shortcuts the moment the two disagree. There's a deeper reading here too — one note argues LLMs operationalize Saussure's *langue*, learning meaning purely as relational structure with no outside referent Can language models learn meaning without engaging the world?. On that view syntactic encoding and semantic representation aren't two systems at all, but two faces of the same relational compression — which is also why, when they conflict, the model has no grounded tiebreaker and defaults to whatever the statistics favor.
Sources 9 notes
The Polar Probe shows LLMs represent syntactic type and direction through both distance and angular position between embeddings, nearly doubling accuracy over distance-only methods. This demonstrates neural networks spontaneously learn structured, symbolic-compatible geometry.
OpenAI's o1 model successfully constructs syntactic trees and phonological generalizations through explicit step-by-step reasoning, revealing that LLM linguistic capability extends far beyond behavioral language tasks to genuine language analysis.
Twenty-eight semantic axes in LLM embeddings reduce to three principal components matching human EPA structure. Intervening on one feature predictably shifts aligned features proportionally, creating unavoidable off-target effects that reflect how meaning is fundamentally organized.
Clustering analysis of RoBERTa embeddings reveals sensitivity to five psycholinguistic measures including valence, concreteness, iconicity, and taboo. This demonstrates that static embeddings function as genuine lexical entries containing semantic content before self-attention operates.
Circuit tracing in Claude models reveals features progress from token-level inputs to abstract concepts to functional operations to outputs. Larger models develop richer abstract features, suggesting scaling enables higher-level conceptual reasoning rather than pattern memorization.
LLMs show consistent preference for higher-frequency surface forms over semantically equivalent rare paraphrases across math, machine translation, commonsense reasoning, and tool calling. This suggests models track statistical mass from pretraining rather than meaning-recognition as their primary mechanism.
When semantic content is decoupled from reasoning tasks, LLM performance collapses even with correct rules in context. Models rely on parametric commonsense and token associations rather than formal logical manipulation, constraining reasoning to training distribution semantics.
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.
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.