Let the Agent Sync: LLM-Driven Code Generation for the Families to Persons Case

Abstract

We present BXAgent, a tool that automatically generates Java code for
bidirectional model transformations between EMF/Ecore metamodels from a
natural-language description. Given two .ecore files and a short
textual specification, an LLM-backed pipeline produces a single Java class
that supports (i) batch transformations in both directions,
(ii) incremental synchronisation after single-sided edits, and
(iii) concurrent synchronisation (csync) when both models are edited
independently. Unlike conventional bx tools that require a formal
specification in a dedicated transformation language (e.g. TGGs or lenses),
BXAgent's only user-facing input is natural language. The generated code is
deterministic and stable: because the transformation logic is produced by a
fixed FreeMarker template, repeated runs yield identical output for the same
mapping specification. We applied BXAgent to the revised Families-to-Persons (F2P)
benchmark of TTC 2026, which extends the classic case with 13 concurrent
test cases and 3 round-trip cases in addition to the original 34 batch and
incremental tests. BXAgent passes 46 out of 50 tests, including all
four genuine conflict cases. We describe the tool architecture,
the generated synchronisation algorithm, our conflict resolution strategy,
and discuss the four failing cases together with their root causes.

Mehr zum Titel