How do hierarchical knowledge layers capture different types of narrative information?
This explores how stacking knowledge in layers — from broad summaries down to specific details — lets a system hold different kinds of narrative information at once, rather than flattening a story into one undifferentiated pile of text.
This explores how layered, hierarchical knowledge structures let a system capture different kinds of narrative information at once — the big-picture arc, the scene-level detail, and the character interiority — instead of treating a story as one flat bag of chunks. The corpus's clearest answer comes from MegaRAG, which builds hierarchical multimodal knowledge graphs over books so that high-level summaries, page-specific details, and even images all live as separate but connected layers; this is exactly what lets it answer cross-chapter, 'global' questions that flat chunk retrieval can't reach Can multimodal knowledge graphs answer questions that flat retrieval cannot?. The hierarchy isn't decoration — it's what makes whole-story reasoning possible.
What's striking is that this coarse-to-fine layering may be a property of how representations naturally organize, not just an engineering choice. The leading eigenvectors of embedding matrices split meaning from broad categories down to fine sub-branches, tracking a taxonomy tree level by level Do embedding eigenvectors organize taxonomy from coarse to fine?. So the 'layers' a knowledge graph imposes by hand echo a structure the model's own geometry already prefers — abstraction at the top, specificity at the bottom.
But different layers capture genuinely different *types* of narrative information, and the corpus shows this from several angles. One layer is structural: AI-versus-human fiction can be told apart using only discourse-level features like character agency and chronological ordering — narrative shape, divorced from sentence style Can AI stories be detected without analyzing writing style?. Another layer is psychological: predicting a character's choices works best when persona profiles are paired with memories retrieved as relevant to that character's mind — a character-state layer distinct from plot summary Can LLMs predict character choices from narrative context?. And a third is temporal: language models segment stories into events at boundaries that match human consensus, carving the continuous flow into discrete units Do language models segment events like human consensus does?. Summary, structure, character, and event-time are not the same information — a good hierarchy keeps them as separable channels.
The deeper payoff of layering is that structure becomes something you can *reason over*, not just retrieve from. Knowledge-graph curricula teach models domain expertise by composing primitives along graph paths rather than relying on scale Can knowledge graphs teach models deep domain expertise?, and symbolic rules derived from graph topology give a model navigational plans that beat plain semantic-similarity search Can symbolic rules from knowledge graphs guide complex reasoning?. Applied to narrative, that's the difference between 'find the chunk that mentions the betrayal' and 'trace how this character's arc connects across six chapters.' The thing you didn't know you wanted to know: the value of hierarchy isn't compression — it's that layers let abstract questions (theme, arc, motivation) find answers that no single passage contains, because the answer lives in the relationships between layers, not in any one of them.
Sources 7 notes
MegaRAG builds hierarchical multimodal knowledge graphs from text and visuals to answer cross-chapter, global questions that flat chunk retrieval cannot reach. The hierarchy supports abstraction levels from high-level summaries to page-specific details while treating images as first-class graph nodes.
Leading eigenvectors of embedding Gram matrices separate broad taxonomic branches first, then progressively finer sub-branches—a coarse-to-fine spectral order that tracks the WordNet hypernym tree level by level, confirming predictions from co-occurrence statistics.
StoryScope achieved 93.2% accuracy separating AI from human fiction using only discourse-level features like character agency and chronological structure, retaining 97% of performance while eliminating stylistic cues. These structural choices resist humanization because they require rewrites, not surface edits.
The LIFECHOICE benchmark (1,462 decisions across 388 novels) shows LLMs predict character choices better when given expert-written persona profiles paired with retrieved memories relevant to the character's psychology. This persona-based approach outperforms automated summarization by 5%.
GPT-3's event boundaries correlate more strongly with averaged human annotations than individual human annotators do. This suggests language models may pre-compute statistical consensus through training on diverse text, or that next-token prediction parallels human event cognition.
Fine-tuning a 32B model on 24,000 reasoning tasks derived from medical knowledge graph paths produces state-of-the-art performance across 15 medical domains, demonstrating that structured knowledge composition matters more than scale.
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.