The Granularity Gap in Hierarchical Summarization

This direction is split into other directions: abstraction-induced information loss. Its last version stays readable and citable here.

Version 1 of 127 papers36 conceptsfrozen 2026-10-04

To process a library, you cannot read every word. The standard solution in natural language processing is to build a hierarchy: summarize individual documents, then summarize those summaries, repeating the process until you reach a single, manageable overview. This recursive abstraction allows models to scale to massive datasets, from multi-document event logs to entire codebases.

However, this structural elegance hides a fundamental flaw known as the aggregation paradox. As information moves up the hierarchy, it undergoes a phase transition from specific data to general themes. In this process, the "granularity"—the names, dates, technical nuances, and specific constraints that make information actionable—is often pruned away.

The Mechanism of Information Decay

Hierarchical summarization works by prioritizing global salience over local detail. Janara Christensen and colleagues noted as early as 2014 that maintaining semantic coherence becomes exponentially harder at deeper levels of a hierarchy. When a system attempts to find a "child" summary that logically follows a highly specific "parent" sentence, the connection often breaks because the parent has already abstracted away the necessary context.

This decay is not just a side effect; it is a built-in trade-off. Nuno Vasconcelos and Andrew Lippman identified this as a tension between computational efficiency and model resolution. By clustering data into coarser abstractions, we gain speed but lose the fine-grained structural details present in the raw source.

The Aggregation Paradox

The most striking evidence of this problem appears in modern Retrieval-Augmented Generation (RAG). Systems like GraphRAG use hierarchical summarization to make sense of large knowledge graphs. Yet, Pengyu Wang and their team discovered that these structured systems often underperform simple "flat" retrieval.

They termed this the aggregation paradox: while the hierarchy helps connect scattered facts, the summarization process itself acts as a filter that discards the very details needed for factual recall. When a user asks a specific question, the high-level thematic nodes in the hierarchy are often too "blurry" to provide a precise answer.

Domain-Specific Failure Modes

The loss of granularity manifests differently across research domains, but the underlying cause remains the same:

  • Event Summarization: Adithya Pratapa and Teruko Mitamura found that hierarchical methods "lose the plot" during final synthesis, omitting critical identifiers like numerical values and specific dates that are present in intermediate stages.
  • Technical Classification: In patent analysis, Benoît Sagot observed that standard importance-ranking algorithms prioritize general themes. This strips away the subtle technical nuances required to distinguish between closely related sub-categories, making automated classification less accurate than using truncated raw text.
  • Medical Evidence: Chantal Shaib and Iain Marshall demonstrated that hierarchical synthesis often fails to capture the "weighing" of evidence. A summary might report a positive outcome while losing the granular caveats about sample size or study bias that a human expert would prioritize.
  • Software Engineering: When summarizing code modules, Weisong Sun found that method-level summaries often lack the nuanced logic needed to construct a coherent file-level overview.

Toward Grounded Abstractions

If the problem is that hierarchies become "ungrounded" as they grow, recent research focuses on keeping high-level nodes tethered to their roots.

Yuntong Hu and colleagues proposed HiGTL, which jointly optimizes structural clustering and semantic labeling to ensure broad category labels remain grounded in fine-grained data points. Similarly, Guanran Luo introduced DTCRS, a framework that builds "query-aware" trees on the fly. Instead of a static hierarchy, the system dynamically decomposes a question to decide which parts of the document need high-level synthesis and which require direct, raw-text access.

The consensus across these 27 papers is that summarization is not a loss-less compression. To build truly effective large-scale systems, researchers must move beyond simple recursive merging and find ways to preserve the "leaf-node" precision that gives data its value. As Joon Sik Kim and Ameet Talwalkar argue, sometimes a summary is actually a hindrance; in tasks like document matching, the "noise" we try to summarize away is often the only signal that matters.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_the_granularity_gap_in_hierarchical_summ_42bbe94e,
  title = {The Granularity Gap in Hierarchical Summarization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-granularity-gap-in-hierarchical-summarization/rdr_42bbe94e66274a038fc42204fffc0ed3},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_the_granularity_gap_in_hierarchical_summ_42bbe94e_v1,
  title = {The Granularity Gap in Hierarchical Summarization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-granularity-gap-in-hierarchical-summarization/rdr_42bbe94e66274a038fc42204fffc0ed3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}