Schema-Derived Decoding Boosts MLIR Code LM Generalization.

Researchers have explored a method to significantly improve the ability of small code language models (LMs) to generate Multi-Level Intermediate Representation (MLIR) code across various dialects without requiring extensive retraining. MLIR is a crucial component of modern machine learning compiler infrastructure, but its extensible nature and limited presence in pretraining data make it challenging for LMs to handle new dialects. The core idea involves using inference-time priors derived directly from each dialect's Operation Definition Specification (ODS). This "schema-derived constraint stack" includes a Control Flow Graph (CFG) over operation signatures, type-domain splits, and an SSA-scope validator. This system allows for robust code generation even for dialects not explicitly seen during training. To validate this, the team released four new natural-language-to-MLIR benchmarks. Experiments showed that a relatively small model, SmolLM2-1.7B, equipped with these schema-derived priors, matched or even surpassed much larger 15B-34B code LMs in terms of verify-valid generation, particularly on structurally constrained dialects like Linalg. This demonstrates a powerful technique for achieving cross-dialect generalization efficiently.