💬 NLP

Attention Flows: Tracing LLM Conceptual Engagement via Story Summaries

This paper evaluates the ability of state-of-the-art LLMs to comprehend long-form narratives by comparing their novel summaries against human-written ones, revealing that models struggle with holistic integration and disproportionately focus on the ends of texts, a finding linked to their attention mechanisms.

Original authors: Rebecca M. M. Hicke, Sil Hamilton, David Mimno, Ross Deans Kristensen-McLachlan

Published 2026-04-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rebecca M. M. Hicke, Sil Hamilton, David Mimno, Ross Deans Kristensen-McLachlan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Can AI Actually "Read" a Whole Book?

Imagine you give a super-smart robot a 500-page novel and ask it to write a summary. In the past, robots were like people with very short attention spans; they could only remember the first few pages and the very last page. If you asked them about the middle, they'd just guess.

Recently, these robots (Large Language Models, or LLMs) have gotten "memory upgrades." They can now hold millions of words in their "mind" at once. The big question this paper asks is: Just because they can hold the whole book, does that mean they actually understand it?

The authors argue that "retrieving" a fact (finding a needle in a haystack) is easy, but "comprehending" a story (understanding the whole plot) is hard. They call the failure to understand the middle of a story "Context Rot." It's like reading a book but your brain starts to rot or decay in the middle chapters, leaving you with a clear memory of the beginning and end, but a fuzzy mess in the middle.

The Experiment: The "Story Map" Test

To test if these robots truly understand stories, the researchers didn't just ask for a summary. They did something clever: They created a "Story Map."

The Human Baseline: They took 150 classic novels (like Dracula or Pride and Prejudice) and looked at the summaries humans wrote on Wikipedia.
The Robot Summaries: They asked 9 different AI models to write summaries of the same books.
The Alignment: This is the tricky part. They took every single sentence from the summaries and asked: "Which specific chapter of the original book does this sentence describe?"

Think of it like this: If a human reads a book, they create a mental map where they visit every room (chapter) to see what happened. The researchers wanted to see if the AI's map looked the same as the human's map.

What They Found: The "End-Weighting" Problem

The results were revealing. While the AI summaries sounded okay, their "mental maps" were very different from humans.

Humans are Tour Guides: When humans summarize a story, they tend to visit every room in the house. They spend a little time in the beginning, a lot in the middle, and a bit at the end. They cover the whole journey.
AI are "Recency Bias" Fans: The AI models acted like they were only interested in the beginning and the end of the book. They largely ignored the middle.
- The Metaphor: Imagine you are describing a road trip to a friend. A human would say, "We drove through the mountains, got lost in the desert, saw a cool canyon, and then arrived at the beach." The AI would say, "We started at the mountains... and then we were at the beach!" It skipped the entire middle of the trip.

The paper calls this "End-Weighting." The AI models seem to think the most important parts of a story are the very first thing that happens and the very last thing that happens, treating the middle chapters as boring filler.

Other Weird Differences

The study also found some other quirks in how AI writes compared to humans:

The "High-Level" Summary: AI summaries were often shorter and more vague. Instead of saying "The hero fought the dragon in the cave," the AI might say "The hero faced a great challenge." It's like a robot giving you the "Cliff's Notes" version without the details.
Less Linear: Humans usually tell stories in order (Chapter 1, then 2, then 3). AI summaries often jumped around, mixing events from different times, as if they were looking at the whole book at once rather than reading it page-by-page.
The "Lost in the Middle" Attention: The researchers looked inside one of the AI models (Qwen 3.5) to see how its "attention" worked. They found that the model literally paid less "attention" to the middle chapters. It was like a flashlight that was bright at the start and end of the book but dim in the middle.

Why Does This Matter?

This is a big deal because we are starting to use AI to analyze huge amounts of text—legal documents, medical records, and history books.

If an AI is good at finding a specific date in a document (the "Needle in a Haystack"), but bad at understanding the story of the document (the "Context Rot"), then we can't fully trust it to make complex decisions. It might miss the crucial plot twist that happens in Chapter 40 of a 500-page contract.

The Takeaway

The paper concludes that while AI has gotten much better at holding long texts in its "memory," it hasn't quite learned how to sustain attention across the whole story. It's like a student who can memorize the first and last page of a textbook perfectly but falls asleep during the middle 400 pages.

The researchers released their data (the 150 books and all the summaries) so other scientists can try to fix this "Context Rot" and teach AI to pay attention to the whole story, not just the beginning and the end.

1. Problem Statement

Despite significant advancements in Large Language Model (LLM) context window sizes (scaling from thousands to millions of tokens), there is growing evidence that models struggle to integrate information across long-form texts. While models often pass "Needle in a Haystack" retrieval tests (locating specific tokens), they exhibit a phenomenon termed "context rot," where they fail to coherently synthesize information throughout a document.

Existing benchmarks often rely on synthetic retrieval tasks or short narrative excerpts, failing to capture the complexity of sustained conceptual engagement required for understanding full-length narratives. The authors argue that current evaluation methods do not adequately measure whether models truly comprehend the narrative structure of long texts or if they merely attend to specific fragments (often the beginning and end).

2. Methodology

The authors propose a novel framework to evaluate conceptual engagement by analyzing how LLMs and humans summarize long-form narratives. The core hypothesis is that a summary acts as a "conceptual map," revealing which parts of a text the author (human or model) deems critical.

Data Collection

Corpus: 150 public domain English novels (ranging from 9,780 to 488,101 tokens, averaging ~147k tokens) sourced from Project Gutenberg.
Human Baseline: Corresponding Wikipedia plot summaries for each novel, co-authored by multiple contributors to minimize individual stylistic bias.
Model Generation: Nine state-of-the-art LLMs (ranging from 7B to 120B parameters, including Qwen, Llama, Gemini, and GPT variants) generated summaries for each novel under four distinct prompt conditions:
1. Full text only.
2. Full text + Wikipedia summarization guidelines.
3. Title/Author only.
4. Title/Author + Guidelines.
- Total Dataset: 5,400 model-generated summaries + 150 human summaries.

Alignment Pipeline

To measure "engagement," the authors developed a sentence-to-chapter alignment task:

Task: Map each sentence in a summary to the specific chapter(s) of the source novel it references.
Validation: Two human experts annotated alignments for four novels, achieving high inter-annotator agreement ( $\kappa = 0.74$ ).
Automation: An automated alignment system was trained and evaluated. GPT-OSS (120b) was selected as the alignment engine, achieving an F1 score of 80% on the validation set, significantly outperforming TF-IDF and embedding baselines.
Metric Construction: The aligned data creates a "conceptual map" allowing for the calculation of:
- Salience Distribution: How many sentences are dedicated to specific chapters.
- Linearity: The degree to which summary sentences follow the chronological order of the source text.
- Skew/End-Weighting: The bias toward the beginning or end of the text.
- Skipping Rates: The proportion of chapters or sentences that are ignored (not matched).

Attention Mechanism Probing

For the open-weight model Qwen 3.5 (9b), the authors analyzed internal attention matrices to correlate the model's internal attention weights with the external summary alignment, testing the "Lost in the Middle" hypothesis.

3. Key Contributions

Conceptual Engagement Framework: Introduces a method to evaluate long-context comprehension by tracing attention flows via summary generation, moving beyond simple token retrieval.
Dataset Release: A large-scale dataset of 5,550 summaries (human and model) aligned at the sentence-chapter level for 150 novels.
Discovery of Structural Bias: Identifies a systematic bias in LLMs where they disproportionately focus on the end of long narratives, a flaw undetectable by standard retrieval benchmarks.
Interpretability Insights: Links the behavioral failure (skipping middle chapters) to internal attention mechanisms, showing that attention mass is not uniformly distributed across the context window.

4. Key Results

Stylistic and Syntactic Differences

Length: Model summaries are generally shorter than human summaries (fewer tokens and sentences), except for the largest models (Llama 4 Scout, GPT 5.4).
Complexity: Models produce summaries with lower syntactic complexity (measured by dependency distance) compared to humans.
Entity Density: Models use fewer named entities, particularly person entities, unless provided with the full text.

Conceptual Engagement Differences

End-Weighting (The "Recency Bias"): This is the most significant finding. While human summaries distribute attention relatively uniformly across the novel (with slight peaks at start, middle, and end), LLMs heavily skew toward the end of the text.
- As model size increases, this end-weighting bias increases when full text is provided.
- Models skip significantly more chapters in the middle of the text compared to humans.
Linearity: Human summaries are generally more linear. Models, especially smaller ones or those without full text, produce less linear summaries, often rearranging events or jumping between chapters.
Hallucination vs. Abstraction:
- Models without full text tend to hallucinate (generate sentences that cannot be aligned to any chapter).
- Models with full text produce "higher-level" summaries, describing broader events but skipping specific details and middle chapters.

Attention Mechanism Correlation

Analysis of Qwen 3.5 (9b) revealed a weak correlation ( $\rho \approx 0.32$ ) between the attention mass a model pays to a chapter and the number of summary sentences dedicated to that chapter.
The model struggles to target attention mass accurately in longer texts; the accuracy of aligning the "most attended" chapter to the summary sentence drops as novel length increases.
Providing summarization instructions improved the model's ability to identify relevant spans (Top-1 accuracy), suggesting priming helps attention heads.

5. Significance and Implications

Beyond Retrieval: The paper demonstrates that "finding" information (retrieval) is distinct from "understanding" it (comprehension). Models can locate tokens but fail to build a coherent mental model of the entire narrative arc.
Evaluation Gap: Current benchmarks (like Needle in a Haystack) are insufficient for evaluating long-context LLMs. The proposed "Attention Flows" metric offers a more nuanced view of how models process long documents.
Model Design: The findings suggest that the "Lost in the Middle" phenomenon is a structural issue in how attention mechanisms distribute focus, not just a data scarcity issue. Future architectures may need to explicitly enforce uniform attention distribution or better integrate middle-context information.
Practical Application: For practitioners using LLMs to analyze long legal, medical, or literary texts, the results warn that models may systematically miss critical information located in the middle of documents, focusing instead on conclusions or introductions.

In conclusion, the paper provides robust evidence that despite massive context windows, LLMs currently lack the sustained conceptual engagement required for true long-form narrative comprehension, exhibiting a distinct bias toward the ends of texts that human readers do not share.