MAGS
A multi-agent automated formalization framework that combines the rigor of formal verification with the automation capabilities of LLM agents. It uses Dafny as a verification-aware intermediate representation to provide machine-checkable safety guarantees for AI-generated code.
Timeline (last 90 days)
MAGS (Multi-agent Auto-formalization Guarantees Safety) is a unified multi-agent framework designed to generate executable programs that simultaneously possess formal safety guarantees.
MAGS uses Dafny as a verification-aware intermediate representation.
The MAGS workflow includes: freezing human-audited APIs and safety requirements as immutable verification baselines, translating LLM-generated code into Dafny, automatically repairing it based on verifier feedback, and compiling back to executable code.
The research team evaluated MAGS on three categories of tasks: 100 CUDA kernels, 100 terminal scripts, and 20 robotic arm tasks, totaling 220 samples.
MAGS achieved a 100% success rate across all 220 samples, generating programs with non-trivial safety guarantees against the frozen specifications.
One failure mode of MAGS occurs when the semantics after auto-formalization fail to fully capture the target behavior.
All Facts (6)
MAGS (Multi-agent Auto-formalization Guarantees Safety) is a unified multi-agent framework designed to generate executable programs that simultaneously possess formal safety guarantees.
50%UnverifiedMAGS uses Dafny as a verification-aware intermediate representation.
50%UnverifiedThe MAGS workflow includes: freezing human-audited APIs and safety requirements as immutable verification baselines, translating LLM-generated code into Dafny, automatically repairing it based on verifier feedback, and compiling back to executable code.
50%UnverifiedThe research team evaluated MAGS on three categories of tasks: 100 CUDA kernels, 100 terminal scripts, and 20 robotic arm tasks, totaling 220 samples.
50%UnverifiedMAGS achieved a 100% success rate across all 220 samples, generating programs with non-trivial safety guarantees against the frozen specifications.
50%UnverifiedOne failure mode of MAGS occurs when the semantics after auto-formalization fail to fully capture the target behavior.
50%