Critic's Eye: Review-Grounded Graph Reasoning for Long-Form Motif Discovery

Xinyu Sun, Zhangyue Yin, Xinyi Ju, Xin WangarXiv 2026 · arXiv

TL;DR

CRITIC’S EYE is a novel graph-based framework designed to discover literary motifs—recurring thematic elements like "blood" or "ambition"—across long-form narratives. By modeling a play as a heterogeneous graph of lines, characters, and scenes, it outperforms massive proprietary models like Gemini-1.5-Pro by nearly 4x, proving that explicit structural reasoning is far more effective than brute-force parameter scaling for deep narrative understanding.

What Problem Does This Paper Address?

Current Large Language Models (LLMs) are excellent at processing local text, but they often struggle with the "big picture" of long-form literature. In narratology, there is a distinction between the fabula (the chronological sequence of events) and the sjužet (the artistic way those events are presented).

Motifs are a key part of the sjužet. A motif like "blood" in Macbeth isn't just a recurring word; its meaning shifts from a symbol of honor in Act 1 to a symbol of guilt in Act 5. Because these instances are spread across thousands of lines and lack consistent wording, standard sequence-based models hit a "semantic ceiling." They suffer from "lost-in-the-middle" effects and fail to connect distant but thematically related moments.

Figure 1: Why motif evidence is spread out. Bottom: the fabula, the play’s events in chronological order. Middle: recurring images, each attaching to events that lie far apart on that order. Top: the themes those images carry.

Methods

The authors propose CRITIC’S EYE, which moves away from treating text as a flat sequence and instead builds a Narrative Topology Graph.

  1. Graph Construction: The model creates a graph with three types of nodes: Atomic Units (individual lines), Agents (characters), and Structural Containers (scenes and acts). It then adds a "Thematic Layer" by using an LLM to extract motif references from centuries of professional literary criticism.
  2. Two-Stage Training:
    • Stage 1: Multi-Task Structural Pre-training. The model learns the "anatomy" of the play through self-supervised tasks: predicting the next line (continuity), identifying which scene a line belongs to (localization), and identifying the speaker (alignment).
  3. Stage 2: Semantic Alignment. The model uses contrastive learning to anchor these structural embeddings to the semantic space of motifs, treating motif discovery as a transductive link prediction task.
Figure 2: The CRITIC’S EYE Graph Construction Pipeline. We fuse the explicit Narrative Backbone (parsed from TEI XML) with a latent Thematic Layer derived from literary criticism. (1) An LLM extracts structured tuples (Reference, Theme, Device) from unstructured commentary. (2) These mentions are semantically clustered into canonical motif nodes. (3) The resulting heterogeneous graph encodes both the linear fabula and the critical sjužet, serving as the substrate for transductive inference. The extraction prompt and output schema are in Appendix B.

Datasets

The researchers introduced MOTIFBENCH, a specialized benchmark centered on Shakespeare’s Macbeth.

  • Narrative Backbone: Parsed from the Folger Shakespeare Library’s XML.
  • Ground Truth: Derived from a "Critical Corpus" of 26 authoritative documents spanning from 1765 to 2020.
  • Complexity: The dataset features a severe "long-tail" distribution, where 90% of thematic concepts appear in three or fewer examples, making it a grueling test for standard machine learning classifiers.

Key Insights and Ideas

  • The "Semantic Ceiling": Even the most powerful proprietary models (GPT-4, Claude 3.5, Gemini 1.5) struggle to exceed ~23% accuracy on this task. This suggests that the bottleneck for deep reading isn't raw computational power, but a lack of topological awareness.
  • Inverse Scaling: Interestingly, the largest open-source models (like GPT-OSS-120B) sometimes performed worse than smaller counterparts because their strong generative priors made them less likely to follow the strict constraints required for motif retrieval.
  • Structural Redundancy: The graph structure provides "shortcuts." If a line's local text is ambiguous, the model can route evidence through the "Scene" node or the "Character" node to find the correct motif. This makes the system much more resilient to noise than sequence-only models.
Figure 4: The Severe Long-Tail Distribution of Lit- erary Motifs. MOTIFBENCH follows a sharp power law (Imbalance Ratio = 35.0), with over 96% of active motifs (187/193) in the sparse “tail” (freq ≤5). This extreme sparsity renders standard random splits ineffec- tive, motivating our transductive formulation to capture these rare signals.

Key Results

CRITIC’S EYE established a new state-of-the-art for motif discovery:

  • Performance: It achieved a 84.8% Hit@5, compared to just 22.6% for the best-performing baseline (Gemini-3-pro-preview).
  • Efficiency: It achieved these results using a lightweight Graph Neural Network (GNN) with a fraction of the parameters of a frontier LLM.
  • Ablation Studies: Removing the structural pre-training caused performance to plummet, confirming that "internalizing the narrative anatomy" is the secret sauce of the model's success.

Limitations and Future Work

  • Genre Constraints: The current model relies on explicit structural markers found in drama (Acts/Scenes). Adapting this to prose fiction (novels) would require complex discourse parsing to identify implicit scene boundaries.
  • The "Well-Studied" Requirement: The framework currently needs high-quality expert commentary to build its thematic layer, limiting its use to canonical works like Shakespeare or Joyce.
  • Closed-World Retrieval: The model selects from a fixed inventory of motifs. Future work could explore "open discovery" where the model identifies entirely new thematic threads ab initio.

The authors envision CRITIC’S EYE not as a replacement for LLMs, but as a "structural adapter" that can help AI move from surface-level processing to deep, human-like thematic reasoning.

Summary updated

Abstract

Deep narrative understanding requires distinguishing what happens from how it is presented, moving beyond the chronological fabula to decode the sjuzet, the discursive organization through which themes are staged. We study this through the recurrence of literary motifs (blood in Macbeth, for instance), where symbolic significance emerges not from local cues, but through repetition-with-variation across the whole work. We operationalize this setting as Computational Motif Discovery, a transductive task that predicts missing line-motif links using the full narrative structure of a play. We propose Critic's Eye, which models the work as a heterogeneous Narrative Topology Graph and performs discriminative inference over global structural evidence. Critic's Eye achieves 84.8% Hit@5 on our benchmark, nearly quadrupling state-of-the-art proprietary foundation models (Gemini-3-pro-preview, ~22.6%). Our analysis reveals that despite massive scaling, sequence-based models hit a "semantic ceiling", struggling to resolve dispersed dependencies. These findings suggest that explicit structural priors are a far more effective inductive bias than parameter scaling for decoding the complex architecture of long-form literature.
BibTeX
@misc{sun2026critic,
  author = {Xinyu Sun and Zhangyue Yin and Xinyi Ju and Xin Wang},
  title = {{Critic's Eye: Review-Grounded Graph Reasoning for Long-Form Motif Discovery}},
  year = {2026},
  eprint = {2610.35011},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.35011},
}
Figure