The authors introduce Object Aligner (OA), an open-source Python library that deterministically scores the similarity between two JSON objects by recursively aligning their structures. The tool addresses the limitations of exact matches and text-based similarity by using the Hungarian algorithm for unordered collections and sequence alignment for ordered ones, awarding partial credit based on schema annotations.
- OA handles complex structured data like graphs or hypergraphs through "referential alignment," which infers a bijection between identifiers to make scores invariant to relabeling.
- The library approximates graph isomorphism using Weisfeiler-Leman color refinement and emits ranked repair suggestions for mismatches at no extra cost.
- When used as a reward inside the GEPA prompt optimizer, Object Aligner helps or stays neutral across all tested datasets.
The authors consider this important because it provides a deterministic, schema-configurable method for evaluating LLM outputs in tasks like information extraction and agentic planning without relying on expensive or opaque LLM judges.