The Granularity of Meaning: Structural Abstraction as the Key to Machine Analogical Reasoning

Large Language Models have demonstrated remarkable capabilities across countless domains, yet they remain stubbornly susceptible to surface level pattern matching when confronted with tasks requiring deep structural understanding. Nowhere is this limitation more apparent than in analogical reasoning involving narratives, where the ability to recognize relational similarities between stories despite divergent surface features marks a cornerstone of human cognition. In their recent work, "Enhancing Structural Mapping with LLM derived Abstractions for Analogical Reasoning in Narratives," Khojasteh et al. present a compelling case for hybrid neuro symbolic architectures, demonstrating that careful abstraction of narrative structure proves more consequential for analogical accuracy than raw model scale.

The challenge of narrative analogy extends beyond simple pattern recognition. When humans engage in analogical reasoning, we map the functional and causal organization of events from one story to another, prioritizing relational structure over lexical overlap. Consider the scenario presented in the paper where a base narrative must be matched to one of two target stories. One target shares surface similarities but diverges in causal structure, a near disanalogy, while the other occupies a different domain entirely but mirrors the underlying relational framework, a far analogy. Human reasoners consistently prioritize the far analogy, recognizing that the structural alignment of events matters more than shared vocabulary or character types. Current end to end LLM systems, however, frequently fail this test, seduced by surface similarities that obscure deeper relational mismatches.

The YARN Framework: Modular Abstraction for Structural Alignment

To address this limitation, the authors propose YARN (YIELDING ABSTRACTIONS FOR REASONING IN NARRATIVES), a modular framework that strategically decomposes the analogical reasoning pipeline into distinct stages. Rather than prompting a monolithic language model to produce an analogy judgment directly from raw text, YARN interposes a critical abstraction layer between unstructured narrative input and structured reasoning mechanisms.

The architecture operates through three sequential components. First, an LLM based decomposition module segments narratives into discrete units representing events or actions. Second, an abstraction module transforms these specific units into generalized representations that capture both the semantic content and the functional role of each unit within its narrative context. Drawing on prior work in framing theory, the authors operationalize four distinct levels of abstraction, ranging from concrete, literal descriptions to highly generalized, role based representations that encode narrative function rather than specific content. Finally, these abstracted units feed into the Structural Mapping Engine (SME), a cognitive architecture derived from Gentner's Structural Mapping Theory (SMT), which performs the alignment between base and target narratives.

This separation of concerns addresses a fundamental limitation in both pure cognitive engines and pure neural approaches. Traditional implementations of SME require pre extracted entities and relations, making them inapplicable to raw textual narratives without extensive manual preprocessing. Conversely, while LLMs can process unstructured text, their performance on analogical tasks proves notoriously sensitive to prompt formatting and vulnerable to surface similarity biases. By using LLMs to generate structured abstractions rather than final judgments, YARN leverages the pattern recognition capabilities of neural models while preserving the rigorous structural alignment mechanisms of symbolic reasoning. The authors have made their implementation publicly available, enabling systematic variation of experimental settings to isolate the contributions of individual components.

The Primacy of Granularity: Experimental Evidence

The experimental results reported in the paper challenge prevailing assumptions about model performance and scale. Across evaluations using the ARN dataset, which contains narrative pairs requiring distinction between analogous and disanalogous relationships, YARN consistently outperformed end to end LLM baselines. More significantly, the authors demonstrate that the specific level of abstraction employed affects performance more profoundly than the size of the language model generating those abstractions.

This finding carries substantial implications for the field. The research reveals that abstractions which are too coarse obliterate critical distinctions necessary for analogical mapping, while excessively fine grained abstractions preserve surface specific details that interfere with structural alignment. The optimal abstraction level captures what the authors term the generalized meaning of units alongside their functional roles within the narrative causal chain, effectively stripping away domain specific details while preserving relational topology.

Error analysis reveals persistent challenges that abstraction alone cannot resolve. The framework struggles with implicit causality, where causal relationships between events remain unstated yet crucial for analogical alignment. Additionally, the authors identify an emerging need for better categorization of analogical patterns in narratives, suggesting that current taxonomies may not adequately capture the full range of structural relationships humans effortlessly recognize. These limitations indicate that while abstraction improves performance significantly, it does not fully bridge the gap between statistical pattern matching and genuine causal reasoning.

Beyond Surface Patterns: