Analogical reasoning is a key driver of human generalization in problem-solving and argumentation. Yet, analogies between narrative structures remain challenging for machines. Cognitive engines for structural mapping are not directly applicable, as they assume pre-extracted entities, whereas LLMs' performance is sensitive to prompt format and the degree of surface similarity between narratives. This gap motivates a key question: What is the impact of enhancing structural mapping with LLM-derived abstractions on their analogical reasoning ability in narratives? To that end, we propose a modular framework named YARN (Yielding Abstractions for Reasoning in Narratives), which uses LLMs to decompose narratives into units, abstract these units, and then passes them to a mapping component that aligns elements across stories to perform analogical reasoning. We define and operationalize four levels of abstraction that capture both the general meaning of units and their roles in the story, grounded in prior work on framing. Our experiments reveal that abstractions consistently improve model performance, resulting in competitive or better performance than end-to-end LLM baselines. Closer error analysis reveals the remaining challenges in abstraction at the right level, in incorporating implicit causality, and an emerging categorization of analogical patterns in narratives. YARN enables systematic variation of experimental settings to analyze component contributions, and to support future work, we make the code for YARN openly available.
Figures & tables
Figure 1: Structural mapping example derived from ARN dataset narratives ( Sourati et al., 2024 ) . The narratives are decomposed into units , which are then converted into abstractions to capture their roles and general meaning, thus facilitating a structural mapping. SB and ST2 form an analogy , whereas SB and ST1 are disanalogous because their final events are opposed to each other. In this example, the system needs to prioritize far analogy over a near disanalogy, i.e., relational over surface similarity.
Figure 2: A high-level overview of the YARN pipeline: we first use LLMs to extract information by identifying units in the two narratives and converting them into abstractions, and then generate a structural one-to-one mapping between the units and abstractions of the two narratives.
Units
Conceptual abstraction
Evaluative abstraction
Narrative arc abstraction
Stage abstraction
Johnny had too many projects
tasks workload
struggle
Main event
work overload
Johnny had too many short deadlines
time pressure
struggle
Main event
Johnny was stressed
emotion stress
struggle
Challenge
stress from workload
Johnny kept working hard
work dedication
effort
Action
persistent effort and completion
Johnny finished everything
tasks completion
gain
Action
Johnny’s boss noticed his hard work
work recognition
gain
Conclusion
reward for dedication
Table 1: The extracted event phrases and abstractions for the following story from the ARN benchmark ( Sourati et al., 2024 ) : “Johnny had too many projects and too many short deadlines, and he was stressed. He kept working as hard as he could to finish everything. His boss noticed how hard he was working and offered him a raise.” For each extracted event, we provide conceptual, evaluative, and narrative arc abstractions. Events that share the same narrative arc (indicated by dashed lines) are then grouped into a single stage abstraction. A detailed explanation of how each abstraction level is extracted is provided in subsection 3.2 .
Figure 3: Story Unit Abstraction. Story events are transformed into abstract representations that capture their underlying functional roles and semantic meaning. By moving beyond surface-level details, these abstractions establish the basis for structural mapping of stories.
Figure 4: Structural Mapping. For each pair of stories, all candidate mappings between units (or their abstractions) are generated and assigned similarity scores. A greedy algorithm is then used to generate a one-to-one mapping, producing a final score that reflects the overall structural correspondence between the stories.
Dataset name
#Target stories
Dataset size
Avg. length (#tokens)
MCQ
4
360
17.3
ARN
2
1,096
65.8
Table 2: Statistics of the MCQ and ARN datasets.
Figure 5: Different levels of conceptual abstraction affect performance depending on the degree of surface similarity. This figure shows the effect of hierarchical conceptual abstraction levels for Qwen, with and without modifiers. Solid bars denote Acon,0 , hatched bars Acon,1 ; pale indigo bars use only the root, and indigo bars use both the modifier and the root. The first pale indigo solid bar in each group corresponds to the setting in Table 3 .
Figure 6: Stage abstraction effectiveness depends on the type of analogy, with lower levels benefiting near cases and higher levels improving far cases. Effect of hierarchical stage abstraction levels with different settings for Qwen. The solid line shows Astg,0 (also reported in Table 3 ), while dashed lines show Astg,1 under three mapping settings: SM (which reduces to cosine similarity since each story is a single phrase), and LLM zero-shot and CoT, where phrases are treated as high-level messages following ( Sourati et al., 2024 ) . Horizontal lines mark random performance for MCQ and ARN.
Figure 7: LLMs’ analogical reasoning is inconsistent across different input units and instructions. The result of using different input and mapping instructions with Qwen for analogical reasoning.
Figure 8: Different patterns appear in the ARN stories when identifying the correct answer. Pattern 1 emphasizes the journey itself, meaning that the order and progression of events are crucial. Pattern 2 focuses mainly on the final outcome, where the sequence of events plays a minor role. Pattern 3 highlights the same underlying rule by presenting opposite or contrasting sequences, where similarity emerges through contrast rather than surface alignment.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A.1: Effect of hierarchical conceptual abstraction levels for Llama, with and without modifiers.
Figure A.2: Effect of hierarchical stage abstraction levels with different settings for Llama.
Figure A.3: The result for using different input and mapping instructions with Llama for analogical reasoning.
Full story
Event phrases
This is the opposite of the base story, so it cannot be the answer.
Although the endings are opposite, the structure is the same, so this is the answer.
Appendix
Table A.1: The model (Qwen) produces different reasoning depending on the input units. In both cases, the model recognizes that the story is the opposite of the base story, but in one case it rejects it for that reason, while in the other it selects it, arguing that the underlying structure is the same. This illustrates the inconsistency of LLMs in analogical reasoning.
Pattern
Story 1
Story 2
P1
Johnny is overwhelmed by work and tight deadlines, but his hard work pays off when his boss recognizes him with a raise and greater choice in future projects.
After a painful breakup and a period of despair, she reluctantly attends a party where she meets someone new, leading to a lasting relationship that is now two years strong.
P2
A former Wi-Fi firmware engineer, forced out of work by illness, turns to low-paid crowdsourced micro work and, after months of struggle, celebrates earning enough in a day to feed his family without draining his remaining savings.
A wireless firmware engineer enters the freelance gig economy to support his family, works diligently through its challenges, and gradually builds a stable and growing income through consistent effort.
P3
A pastor realizes too late that focusing on preaching to others has come at the cost of neglecting his own children, who grew up without his guidance and became unbelievers.
A devout father regularly brings his children to worship and spiritual activities, passing on his faith and values and teaching them that devotion and helping others begin at home.
Appendix
Table A.2: Illustrating examples of different patterns that lead to the correct analogy in the ARN dataset. P1 : Arc-based similarity; P2 : Outcome-based similarity; P3 : Contrastive similarity.
## Role Assignment
You are an annotator who reads short stories and converts them into a list of event phrases.
—
## Task Definition
For each story, extract every distinct event and express it as a single, concise phrase following the rules below.
—
## Term Definitions
Appendix
Table B.1: Prompt for Event extraction
## Role Assignment
You are a frame-extraction assistant. Your job is to assign exactly one concise semantic Frame to each phrase, and must follow the schema [MODIFIER]_[ROOT].
—
### Task Definition
Use the following specification to construct each Frame:
ROOT*: Purpose: Name the core event/state/concept in one noun-like word.
MODIFIER*: Purpose: Provide an explanation for the ROOT in one word.
Analogical reasoning studies how systems transfer relations and solutions between situations with shared structure, enabling inference beyond direct examples.
Analogical reasoning is a core cognitive faculty essential for narrative understanding. While LLMs perform well when surface and structural cues align, they struggle in cases where an analogy is not apparent on the surface but requires latent information, suggesting limitations in abstraction and generalisation. In this paper we compare a model's probed representations with its prompted performance at detecting narrative analogies, revealing an asymmetry: for rhetorical analogies, probing significantly outperforms prompting in open-source models, while for narrative analogies, they achieve a similar (low) performance. This suggests that the relationship between internal representations and prompted behavior is task-dependent and may reflect limitations in how prompting accesses available information.
Hope McGovern, Caroline Craig, Thomas Lippincott +1
Cambridge University · Northeastern University · Athenahealth +1
While LLMs demonstrate remarkable fluency in narrative generation, existing methods struggle to maintain global narrative coherence, contextual logical consistency, and smooth character development, often producing monotonous scripts with structural fractures. To this end, we introduce PLOTTER, a framework that performs narrative planning on structural graph representations instead of the direct sequential text representations used in existing work. Specifically, PLOTTER executes the Evaluate-Plan-Revise cycle on the event graph and character graph. By diagnosing and repairing issues of the graph topology under rigorous logical constraints, the model optimizes the causality and narrative skeleton before complete context generation. Experiments demonstrate that PLOTTER significantly outperforms representative baselines across diverse narrative scenarios. These findings verify that planning narratives on structural graph representations-rather than directly on text-is crucial to enhance the long context reasoning of LLMs in complex narrative generation.
Hanwen Gu, Chao Guo, Junle Wang +2
Institute of Automation, Chinese Academy of Sciences · Tencent Turing Lab
Systematic comparisons between current situations and structurally similar past events in the historical, i.e., historical analogies, is among the most powerful tools for foresight analysis. In this work, we present a new task called Analogical Deep Research (ADR) to Large Language Model (LLM) agents and construct the first ADR benchmark ADR-bench to study whether LLM agents are able to find and leverage historical analogies when doing foresight analysis. Our investigation reveals a key obstacle: LLM agents are poor at finding analogies because they match on surface features rather than underlying mechanisms. We argue that ADR is inherently a causal question as it requires understanding why the event occurred. Based on our theoretical analysis, we propose two principles required for ADR, including the mechanism alignment and cross-analogy confirmation. Built upon our theoretical results, we propose a new agentic framework called Causal Analogical Researcher (CANA) that guides LLMs to find and integrate historical analogies. CANA incorporates a simple yet effective structural decomposition representation, and integrates structural feedback for reflective improvements of historical analogy identification and integration. We show that CANA brings up to 10% improvements in historical analogy generation, and surpasses the state-of-the-art deep research agents in the ADR-bench. Case studies with the ongoing events confirm the effectiveness of CANA in leveraging historical analogies.
Yongqiang Chen, Guangyi Chen, Yuewen Sun +1
1Mohamed bin Zayed University of Artificial Intelligence · 2Carnegie Mellon University