Can language models perform purely symbolic reasoning when semantics are removed?
This explores whether LLMs can do formal logic on its own — manipulating rules and symbols by their structure alone — once you strip away the real-world meaning of the words, or whether their reasoning is fundamentally tied to the semantics of the content.
This explores whether LLMs can do formal logic on its own — pushing symbols around by their structure once the real-world meaning is stripped away. The corpus answer is largely no, but with a fascinating wrinkle: the machinery for symbolic reasoning seems to exist, it just can't operate cleanly when meaning is removed. The most direct evidence is that when semantic content is decoupled from a reasoning task, LLM performance collapses — even when the correct rules are handed to the model in context Do large language models reason symbolically or semantically?. The models lean on parametric commonsense and token associations rather than manipulating logical form, so reasoning stays trapped inside the semantics of the training distribution.
But look one layer down and the picture gets more interesting. Inside the network there actually is a content-independent reasoning circuit — a three-stage mechanism (recitation, middle-term suppression, mediation) that runs across architectures regardless of the specific content How do language models perform syllogistic reasoning internally?. So the symbolic skeleton is real. The problem is that separate attention heads carrying world knowledge keep contaminating it, dragging conclusions toward what's *plausible* rather than what's *logically valid* — and that contamination gets worse at larger scale, not better. In other words, semantics isn't just a crutch the model could drop; it actively overrides the formal machinery.
This reframes "failure" in a useful way. When reasoning models break down, it's often not the logic that fails but the bookkeeping: models confined to text generation can't execute long multi-step procedures even when they demonstrably know the algorithm, and giving them tools lets them sail past the supposed reasoning cliff Are reasoning model collapses really failures of reasoning?. Relatedly, breakdowns track *instance novelty*, not task complexity — models fit patterns tied to specific instances they've seen rather than learning a generalizable, content-blind algorithm Do language models fail at reasoning due to complexity or novelty?. Both point to the same thing: what looks like symbolic competence is often memorized semantic pattern-matching wearing a logical costume.
The most practical thread in the collection suggests the right move isn't to remove semantics at all. Partial formalization — enriching natural language with selective symbolic structure rather than replacing it — beats *both* pure language and full symbolic translation, with full formalization actually losing information that the model needs Why does partial formalization outperform full symbolic logic?. And when you prune reasoning chains down to what matters, the model preferentially protects the symbolic-computation tokens while discarding grammar and filler Which tokens in reasoning chains actually matter most? — evidence that the model itself treats symbolic operations as load-bearing.
The thing you might not have known you wanted to know: the question "can it reason symbolically without semantics?" may be the wrong test entirely. These models appear to be hybrid engines where a genuine logical circuit and a semantic-association engine run on the same hardware and interfere with each other. Stripping semantics doesn't reveal a clean symbolic reasoner underneath — it removes the very signal the model uses to know which operations to run.
Sources 6 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.
Models confined to text-only generation cannot execute multi-step procedures at scale, even when they know the underlying algorithm. Tool-enabled models solve problems beyond the supposed reasoning cliff, suggesting the bottleneck is procedural execution bandwidth.
LRMs don't break at complexity thresholds but at instance-novelty boundaries. Models fit instance-based patterns rather than generalizable algorithms, so any reasoning chain succeeds if trained on similar instances, regardless of length.
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.
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.