microsoft/monitors4codegen

Provides a Monitor-Guided Decoding framework that uses Language Server Protocol clients (via the `multilspy` library) to apply static analysis constraints during LM inference, enforcing type safety and semantic validity across Java, Rust, C#, and Python. The approach intercepts token generation to filter invalid identifiers and method calls by querying live language servers for repository-wide type information, achieving 19-25% compilation rate improvements without model retraining. Includes the PragmaticCode dataset (100 real-world repos with full build environments) and DotPrompts benchmark (10k+ method-completion examples) enabling evaluation on compilation success alongside traditional metrics.

280 stars. No commits in the last 6 months.