Enhancing Structural Mapping with LLM-derived Abstractions for Analogical Reasoning in Narratives

2026-03-31 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new modular framework, YARN (Yielding Abstractions for Reasoning in Narratives), enhances machines' analogical reasoning in narratives by integrating Large Language Models (LLMs) with structural mapping. Analogical reasoning, crucial for human generalization, has been difficult for machines due to the need for pre-extracted entities in cognitive engines and LLMs' sensitivity to prompt format and surface similarity. YARN addresses this by using LLMs to decompose narratives into units, abstracting these units at four defined levels, and then feeding them to a mapping component for cross-story alignment. Experiments demonstrate that these LLM-derived abstractions consistently improve model performance, achieving competitive or superior results compared to end-to-end LLM baselines. Remaining challenges include optimal abstraction levels and incorporating implicit causality.

Key takeaway

For research scientists developing AI systems for complex reasoning, YARN offers a validated approach to improve analogical capabilities in narrative understanding. You should consider integrating LLM-derived abstractions into your structural mapping pipelines to overcome limitations of surface similarity and enhance generalization, especially when dealing with diverse narrative structures.

Key insights

LLM-derived narrative abstractions significantly improve machine analogical reasoning by enhancing structural mapping.

Principles

• Abstraction improves analogical mapping.
• Narrative decomposition aids reasoning.

Method

YARN uses LLMs to decompose narratives into units, abstract these units at four levels (general meaning, story role), and then maps elements across stories for analogical reasoning.

In practice

• Decompose narratives into abstract units.
• Define abstraction levels for story roles.

Topics

• Analogical Reasoning
• Narrative Structures
• LLM Abstractions
• Structural Mapping
• YARN Framework

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning — Human and LLM everyday reasoning aligns with content-sensitive pattern-matching, challenging abstract world model assumptions.
• Structured Abductive-Deductive-Inductive Reasoning for LLMs via Algebraic Invariants — A symbolic scaffold improves LLM logical reasoning by separating inference types and enforcing consistency via algebraic invariants.
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents — Memory retrieval for LLM agents is more effective as an active, reconstructive process over a graph.
• StoryCoder: Narrative Reformulation for Structured Reasoning in LLM Code Generation — Narrative reformulation of code problems significantly improves LLM code generation by structuring reasoning.
• YaRN: Extending RoPE Without Breaking It — YaRN (Yet another RoPE extensioN) is a highly efficient method for extending the context window of pre-trained Large Language Models, such as LLaMA 2, from 4,096 to 128,000 tokens with significantly r…
• LLM-Based Visual Explanation Evaluation Framework for Assessing the Explainability of Facial Skin Disease Classification Models — An LLM-based framework evaluates Grad-CAM explanations for facial skin disease models, ensuring clinical relevance and trustworthiness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.