What question did this study set out to answer?

This research aims to enhance game-engine code generation through automated preference signaling using compiler diagnostics.

June 3, 2026Open Access

Compile-Gated DPO: Executable Verification as an Automated Preference Signal for Game-Engine Code Generation

Key Points

This research aims to enhance game-engine code generation through automated preference signaling using compiler diagnostics.
Developed Compile-Gated DPO to create DPO training pairs from C# completions.
Implemented a fallback system utilizing Unity dotnet, UnityEngineShim, and static analysis for CI portability.
Persisted compile results in DuckDB for statistical comparison across runs.
Successfully generated error-conditioned rejection pairs that improved model learning efficiency.
Showed structured error diagnostics provided valuable context for the model's decision-making process.
Enabled the first use of game-engine compiler diagnostics as effective DPO rejection signals.

Abstract

We present Compile-Gated DPO (CG-DPO), which automatically generates DPO training pairs by writing model-generated C# completions to a scratch namespace, triggering Unity's incremental compiler, capturing structured error diagnostics, and emitting (prompt, chosen, rejected) pairs where the rejection context includes the full diagnostic message. Error-conditioned rejection examples teach the model not just what failed but why. CG-DPO implements a three-tier fallback (Unity dotnet, UnityEngineShim, static analysis) for CI portability. Compile results persist to DuckDB for run-over-run statistical comparison. The first system to use game-engine compiler diagnostics as error-conditioned DPO rejection signals.

Compile-Gated DPO: Executable Verification as an Automated Preference Signal for Game-Engine Code Generation

Key Points

Abstract

Cite This Study