Can a single meeting summary format serve both scanning and reference needs?
This explores whether one meeting-summary layout can satisfy two different reading modes at once — fast scanning ('what do I need to know?') and detailed reference ('what exactly was said, and by whom?') — and the corpus suggests the answer is no, because those modes pull a summary's structure in opposite directions.
This explores whether one meeting-summary layout can serve both quick scanning and detailed lookup. The most direct evidence says it can't. A user study of meeting summaries found three failures, and one of them was exactly this: a single format cannot serve both quick scanning and detailed reference needs at the same time Why do LLM meeting summaries fail to help individuals?. The same study surfaces *why* this is hard — the systems summarize what mattered to the group as a whole, not what mattered to you, and they mis-attribute who said what. Scanning wants the group-level gist compressed; reference wants speaker-accurate, individually-relevant detail preserved. Those are different documents, not two views of one.
The more interesting clue is that the tension isn't just about length — it's about how a summary is *organized*. Work comparing AI and human writing found that ChatGPT defaults to anaphoric organization (recapping what was already said) while human writers lean cataphoric (pointing forward, previewing what's coming) Does ChatGPT organize text differently than human writers?. Scanning is served by forward-pointing structure — headers and previews that let you decide what to read. Reference is served by the backward-looking recap. A model that defaults to one mode is structurally biased against the other, which is part of why a single auto-generated format feels wrong for one of the two jobs.
Where the corpus hints at a way out is in systems that abandon the single flat artifact for a *hierarchy of abstraction*. Knowledge-graph approaches over long documents deliberately support multiple zoom levels — from high-level summaries down to page-specific detail — so the same underlying structure answers both global and pinpoint questions Can multimodal knowledge graphs answer questions that flat retrieval cannot?. The same instinct shows up in retrieval: building a global summary first, then conditioning lookup on it, lets a system connect scattered evidence by its role in the whole rather than by surface similarity Can building a document map first improve retrieval over long texts?. Translated to meetings: the resolution isn't one format, it's one layered structure where scanning reads the top and reference drills down.
There's also a grounding lesson worth borrowing from an unrelated corner — recommendation identifiers. The finding there is that no single representation balances every need: combining IDs, titles, and attributes solves distinctiveness, semantics, and grounding *simultaneously* in a way none of them does alone Can item identifiers balance uniqueness and semantic meaning?. The meeting-summary version of that lesson is that scannability and reference-accuracy are competing facets you compose, not collapse.
So the honest answer: a single format can't do both well, and the field's own work on long-document understanding suggests the fix isn't a cleverer one-size summary but a multi-level structure — a fast top layer for scanning that links down into speaker-attributed detail for reference. The thing you didn't know you wanted to know: the same hierarchy-over-flat-text idea that makes retrieval work on books is exactly what would let one meeting artifact serve two readers.
Sources 5 notes
A user study of seven participants found three critical failures: systems summarize global importance rather than individual relevance, mis-attributions damage group trust and accountability, and one format cannot serve both quick scanning and detailed reference needs.
ChatGPT defaults to summarizing what was already said, while students use more forward-pointing structure that previews upcoming arguments. This reflects different reader models and may stem from how autoregressive generation works token by token.
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.
MiA-RAG inverts standard RAG by summarizing documents first, then conditioning retrieval on that global view. This approach recovers discourse structure that bag-of-chunks retrieval destroys, making scattered evidence findable by their document role rather than surface similarity alone.
TransRec shows that combining numeric IDs, titles, and attributes into structured identifiers solves three problems simultaneously: distinctiveness from IDs, semantics from text, and generation grounding from structural constraints. Neither pure IDs nor pure text alone achieves all three.