Context Graphs: Free Association, Consistency, Narratives & Truth

2025-11-17 · Source: The Ontologist · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, long

Summary

Large Language Models (LLMs) are best understood through the analogy of "free association" or "daydreaming," operating as semi-random diffusion across a graph guided by contextual constraints. Unlike traditional knowledge graphs (KGs), LLMs lack a persistent memory or formal world model, functioning primarily as components within larger systems like Retrieval Augmented Generation (RAG). The article explains how narratives, which constitute most real-world information, can be serialized into sentence graphs with weighted tokens, forming narrative fractals. It demonstrates how a prompt can generate a narrative tree and how LLMs process these by vectorizing tokens, finding aligned threads, and matching subgraphs. Key differences between LLMs and KGs lie in consistency and truth; KGs offer deterministic consistency, while LLMs introduce stochastic noise (temperature) to broaden conversational matches, leading to "hallucinations" as a trade-off for richer output. Both LLMs and KGs are epistolic, reflecting interpreted knowledge rather than absolute truth, with KGs generally offering better curation and consistency.

Key takeaway

For AI Architects and Research Scientists designing knowledge systems, recognize that LLMs excel at free association and narrative generation but inherently lack the deterministic consistency of knowledge graphs. You should consider LLMs as powerful, stochastic components for exploring information, while relying on knowledge graphs for auditable, consistent data storage and structured world models. Balance the "temperature" parameter in LLMs to manage the trade-off between creative output and potential hallucinations, and integrate them with KGs to mitigate consistency issues for critical applications.

Key insights

LLMs function via free association on narrative graphs, trading consistency for broader, stochastically generated outputs.

Principles

• Information is fractal and often narrative-based.
• LLMs are components, not standalone sentient agents.
• Consistency is distinct from truth in knowledge systems.

Method

Narratives can be represented as weighted sentence graphs, where tokens form connections. LLMs process prompts by vectorizing tokens, aligning them with existing subgraphs, and generating responses through a free association process influenced by a "temperature" parameter.

In practice

• Use LLMs for refining ideas or exploring information spaces.
• Adjust LLM "temperature" to balance output richness and relevance.
• Integrate LLMs with KGs for enhanced consistency and memory.

Topics

• Large Language Models
• Knowledge Graphs
• Narrative Representation
• Causality and Influence
• Data Consistency

Best for: AI Architect, Research Scientist, AI Researcher, AI Scientist, Data Scientist

Related on AIssential

• The Internet of Probability — LLMs transform the internet into a probability machine, eroding truth and provenance through statistical generation and contaminated training data.
• The slain cliché: how one random word can kill a story — and resurrect your model’s hidden map — Concept seeding reveals LLM internal knowledge maps and claim stability by perturbing prompts with single, geometrically sampled words.
• LLMs+Graphs: Toward Graph-Native, Synergistic AI Systems — Integrating LLMs with graph computation creates synergistic AI systems that overcome LLM reasoning limitations.
• Operads for compositional reasoning in LLMs — Operads provide a rigorous mathematical framework for question decomposition in LLMs, enabling new consistency metrics for multi-step reasoning.
• HOLONS: A New Hope — Context graphs, structured holonically, enable robust, queryable representations of dynamic information without requiring a triple store.
• Navigating Unreliable Parametric and Contextual Knowledge: Explicit Knowledge Conflict Resolution for LLM Inference — MACR resolves LLM knowledge conflicts by explicitly assessing confidence in sources and employing multi-agent reasoning, moving beyond binary reliability assumptions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Ontologist.