How Little is Enough? Discovering Minimal Prompts through Analogical Abstraction for LLM Reasoning

ACL ARR 2026 January Submission9584 Authors

Keywords: Analogical abstraction, structural sufficiency, minimal prompts, prompt compression, large language models, analogical reasoning

Abstract: Large language models (LLMs) often solve complex tasks with prompts that far exceed the information strictly required for correct reasoning. This raises a fundamental question: what is the minimal structural description sufficient to elicit correct reasoning in an LLM? We propose a diagnostic framework to empirically probe this question using analogical abstraction. Rather than treating prompt compression as an efficiency problem, we view it as a tool for measuring the structural sufficiency threshold of model reasoning. Using code generation tasks with executable unit tests, we iteratively replace original problem descriptions with progressively abstracted analogical descriptions, testing whether functional equivalence is preserved. This allows us to identify the point at which removing structural information causes reasoning failure. Our results show that (i) LLMs can often reconstruct correct solutions from highly abstract analogical descriptions, (ii) different abstraction styles exhibit distinct failure modes, and (iii) models vary substantially in their sensitivity to structural information. These findings suggest that analogical abstraction provides a principled probe for analyzing the minimal conditions under which LLMs retain structured reasoning, offering insights into how pretrained models internalize and retrieve task structure.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: probing, robustness, counterfactual/contrastive explanations, knowledge tracing/discovering/inducing, hardness of samples, free-text/natural language explanations

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis

Languages Studied: English

Submission Number: 9584