Reproduce “JSON Whisperer: Efficient JSON Editing with LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sarel Duanis, Asnat Greenstein-Messica, Eliya Habba
A framework from Lightricks enables Large Language Models to efficiently edit JSON documents by generating minimal RFC 6902 diff patches and employing Explicitly Addressed Sequence Encoding (EASE) for robust array handling. This method matches the quality of full document regeneration while reducing token usage by 31% and cutting processing time and cost by over 30%.
Key insights
- Transitioning from full JSON regeneration to generating minimal diff patches substantially improves efficiency for LLM-based editing while maintaining comparable quality.
- LLMs' inherent difficulties with array index arithmetic and related updates can be effectively resolved through an order-independent encoding like EASE.
- Optimizing few-shot learning with high-quality, synthetically generated examples is essential for LLMs to master the nuances of RFC 6902 patch generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.