2026 Project

SafeCodeRL: Safety-Constrained LLM Code Generation

A multi-agent reinforcement learning framework for dynamic security constraints in LLM code generation.

AI SafetyMulti-Agent SystemsConstrained Reinforcement LearningSecure Code GenerationLLMIoT/CPS SecurityTrustworthy AISensor NetworksEdge ComputingVulnerability Mitigation

SafeCodeRL studies how to reduce vulnerable code generation while maintaining functional correctness.

Publication status: Published on June 2, 2026.

Paper labels: SCI Zone 3; CCF C.

SafeCodeRL framework

The project proposes a collaborative workflow across multiple agents and introduces a constraint-aware policy optimization component. The manuscript reports a substantial reduction in high-risk vulnerable outputs while preserving the utility of generated code.

Publication notes:

  • Published on June 2, 2026.
  • Venue-specific details can be added when a stable public link is available.
  • The project page emphasizes method, responsibility, and reproducible safety evaluation.