Abstract
<title>Abstract</title> <p>Faithfully converting unstructured text into structured representations remains a fundamental challenge. Existing approaches either generate free-text summaries, which are prone to hallucination and omission, or populate fixed, manually engineered schemas that generalize poorly across heterogeneous documentation styles. In both cases, the target representation is fixed in advance, so defi- ciencies in the representation propagate to every extraction. In this work, we present an iterative feedback–refinement framework for adaptive schema induc- tion using large language models (LLMs), which makes the schema itself the object of optimization. A dual-LLM loop iteratively evaluates the schema against source documents and incorporates aggregated feedback into successive schema revisions, progressively improving a shared schema without labeled data or task- specific fine-tuning. To assess reliability, we adopt a faithfulness measure that decomposes quality into precision (absence of hallucination) and recall (absence of omission). Experiments on the public MTSamples clinical corpus with 8 open-weight LLMs compare the framework against extractive, abstractive, and in-context-learning baselines. Iterative schema refinement achieves the highest mean faithfulness across all three LLM-judge metrics while reducing document length to 34.5% of the source. The improvement is driven primarily by higher recall while maintaining comparable or higher precision, and mid-sized models match or outperform larger ones. These findings indicate that optimizing the schema, rather than individual outputs, provides an effective, model-agnostic approach for faithful structured representation.</p>