How does in-context semantic reasoning differ from symbolic reasoning in concept fusion?
This explores the divide between how LLMs actually combine ideas — through meaning-based association (semantic) versus formal rule-following (symbolic) — and what the corpus says about why that distinction matters when models fuse concepts together.
This explores the difference between two ways a model can put concepts together: by leaning on what words *mean* and tend to go with (semantic association), or by manipulating ideas through formal rules regardless of their content (symbolic logic). The most direct answer in the collection is that LLMs are fundamentally semantic, not symbolic. When researchers strip the familiar meaning out of a reasoning task — keeping the logical structure intact but swapping in nonsense or unexpected content — model performance collapses, even when the correct rule is sitting right there in the prompt Do large language models reason symbolically or semantically?. The implication is that models aren't fusing concepts by applying logic; they're fusing them by riding the associations baked into their training distribution.
You can see the seam between the two modes inside the model itself. Studies of how LLMs handle syllogisms find a content-independent circuit — a three-stage mechanism that would, in principle, work like formal logic — but it gets *contaminated* by separate attention heads carrying world knowledge, which tug conclusions toward whatever is semantically plausible rather than logically valid How do language models perform syllogistic reasoning internally?. So the symbolic machinery exists, but the semantic machinery overrides it, and the contamination gets *worse* at larger scale, not better. Concept fusion, in this light, is semantics quietly hijacking a logic engine.
The interesting twist is that pure symbolism isn't the fix. When researchers fully formalize a problem into clean logical notation, they actually lose information — the messy semantic content that the model needs. The sweet spot is *partial* abstraction: enrich natural language with selective symbolic scaffolding rather than replacing meaning with rules, which yields measurable accuracy gains where neither pure language nor full formalization does Why does partial formalization outperform full symbolic logic?. A related approach pulls explicit symbolic rules out of knowledge-graph *structure* to give reasoning a navigational backbone that semantic similarity alone can't provide Can symbolic rules from knowledge graphs guide complex reasoning?. Both say the same thing from opposite directions: semantics supplies the content, symbols supply the structure, and fusing concepts well means keeping both.
This tension shows up even at the token level. When you prune a reasoning chain down to what the model treats as essential, the *symbolic computation* tokens are preferentially preserved while grammar and filler get dropped first — suggesting the model does internally privilege a symbolic skeleton, even if it reasons semantically around it Which tokens in reasoning chains actually matter most?. And in open-ended agentic reasoning over graphs, what keeps fresh concept-combinations coming is that *semantic* surprise persistently outweighs *structural* connection — roughly 12% of links stay semantically novel even after the structure is settled, and that gap is the engine of discovery Why do reasoning systems keep discovering new connections?.
The thing you might not have expected to want to know: there's a whole line of work trying to move the fusion *out* of token-space entirely. Meta's Large Concept Model reasons over sentence-level embeddings in a language-agnostic space before decoding back to words — treating concepts as the unit of computation rather than tokens Can reasoning happen at the sentence level instead of tokens?. That's a third path between raw semantics and rigid symbols: fuse concepts as geometric objects in an abstract space. So 'semantic vs. symbolic' isn't really a binary — it's a spectrum the field is actively trying to find the right point on.
Sources 7 notes
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.
LLMs implement a content-independent three-stage reasoning mechanism—recitation, middle-term suppression, mediation—that works across architectures. However, additional attention heads encoding world knowledge systematically bias conclusions toward semantically plausible rather than logically valid answers, with contamination increasing at larger scales.
QuaSAR and Logic-of-Thought both achieve 4-8% accuracy gains by enriching natural language with selective symbolic elements rather than replacing it. Full formalization loses semantic information; pure language lacks structure. Augmentation preserves both.
SymAgent derives symbolic rules from KG structure using LLM reasoning to create navigational plans that align natural language with graph topology. This approach captures structural reasoning patterns explicitly, outperforming retrieval methods that rely on semantic similarity alone.
Greedy likelihood-preserving pruning reveals six functional token categories; symbolic computation tokens are preferentially preserved while grammar and meta-discourse are pruned first. Student models trained on these pruned chains outperform those trained on frontier-model compression.
Analysis shows iterative graph reasoning evolves toward a stable phase where semantic entropy persistently dominates structural entropy, with ~12% of edges remaining semantically surprising despite structural connection, fueling ongoing discovery.
Meta's Large Concept Model operates on sentence embeddings rather than tokens, reasoning in a language-agnostic space before decoding to any target language. This hierarchical approach with paragraph-level planning produces more coherent output than flat token generation.