Token-Efficient Data Reasoning Agents via Adaptive Structuring of Unstructured Data
The paper’s core claim is a 53% cost cut by caching structure discovered during agent reasoning.
The authors propose “agentic data cracking,” where a sub-agent extracts reusable structured facts whenever the main agent has already opened a document.
The method is meant to avoid repeatedly spending large context windows on the same unstructured sources.
On FanOutQA with one related question added per test question, they report lower cost while preserving accuracy.
ArXiv · AI/CL/LG's note
The authors propose “agentic data cracking,” where a sub-agent extracts reusable structured facts whenever the main agent has already opened a document.
The method is meant to avoid repeatedly spending large context windows on the same unstructured sources.
On FanOutQA with one related question added per test question, they report lower cost while preserving accuracy.
ArXiv · AI/CL/LG's note
score 5