Reification Gap in LLM Modeling

The gap between what language models manipulate (statistical patterns in text) and what they appear to represent (concepts, facts, reasoning). LLMs work with token sequences, not the things those tokens refer to.

The Gap

When we ask an LLM about “Paris,” it doesn’t access some internal representation of Paris-the-city. It predicts tokens based on patterns of how “Paris” co-occurs with other tokens in training data.

Implications

Brittle Grounding: Model “knowledge” is surface patterns, not grounded understanding Sensitivity to Framing: Same question phrased differently gets different answers Hallucination Origin: Models generate plausible-sounding sequences without verifying reference Limits of Scaling: More data improves pattern coverage, not referential grounding

Why It Matters

The reification gap explains why LLMs can be simultaneously impressive (fluent, knowledgeable-seeming) and unreliable (confident confabulation). They’re doing something different from what their outputs suggest.

Mitigation

• Ground LLM outputs in external knowledge sources (RAG)
• Use verification steps before consequential actions
• Don’t treat LLM outputs as authoritative references

Related: 05-atom—uniform-confidence-problem, 05-atom—hallucination-inherent