Runtime Execution Traces Guided Automated Program Repair with Multi-Agent Debate
New AI framework uses runtime snapshots and agent debate to fix complex logic errors in code.
A team of researchers has introduced TraceRepair, a novel multi-agent framework designed to tackle one of software engineering's toughest challenges: automatically fixing complex logic bugs that evade current AI tools. Traditional LLM-based Automated Program Repair (APR) methods work statically, analyzing only source code and test outputs. This approach often misses subtle runtime behaviors and dynamic data dependencies, leading to patches that pass tests by coincidence but don't correct the underlying logic. TraceRepair fundamentally shifts this paradigm by integrating concrete runtime execution traces as objective constraints to guide and validate the repair process.
The system's architecture features specialized AI agents working in concert. A 'probe agent' captures execution snapshots of critical program variables, creating a factual basis for repair. A committee of other agents then uses these runtime facts to debate, cross-verify, and iteratively refine candidate code patches. This multi-agent debate exposes logical inconsistencies that a single LLM might overlook, ensuring fixes are logically sound, not just syntactically plausible. Evaluated on the standard Defects4J benchmark, TraceRepair correctly fixed 392 defects, a substantial leap over previous LLM-based APR methods.
The research demonstrates that the performance gains come from dynamic reasoning, not dataset memorization, as shown by strong results on a new dataset of recent bugs. By treating runtime evidence as a shared constraint for validation rather than mere additional context, TraceRepair provides a more robust, generalizable path toward reliable automated software maintenance. This represents a significant step forward in making AI a dependable partner for developers debugging complex, real-world systems.
- Uses runtime execution traces as objective constraints, not just input, to prevent overfitting and coincidental fixes.
- Employs a multi-agent system with a probe agent for snapshots and a debating committee for patch validation.
- Achieved state-of-the-art results, correctly fixing 392 defects on the Defects4J benchmark, outperforming existing LLM-based methods.
Why It Matters
Moves AI bug-fixing from syntactic guesswork to logic-based reasoning, promising more reliable automated maintenance for complex software.