Reproduce “PoC-Adapt: Semantic-Aware Automated Vulnerability Reproduction with LLM Multi-Agents and Reinforcement Learning-Driven Adaptive Policy” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Phan The Duy, Khoa Ngo-Khanh, Nguyen Huu Quyen, Van-Hau Pham

PoC-Adapt is an end-to-end framework for automated Proof-of-Concept exploit synthesis and verification, combining a multi-agent LLM system with a semantic validation oracle and a reinforcement learning-driven adaptive policy. The system achieved a 15% success rate on real-world vulnerabilities, outperforming a baseline by 25% and reducing the average time-to-exploit steps by 54%.

Key insights

  • Semantic validation, by rigorously comparing pre/post-exploit system states, is crucial for overcoming LLM hallucination and ensuring reliable automated exploit verification beyond superficial execution signals.
  • Reinforcement learning, specifically a DDQN-driven adaptive policy, can effectively optimize an LLM's action space in exploit generation, significantly improving efficiency and reducing token consumption by learning from past attempts.
  • A modular multi-agent architecture with specialized roles and structured feedback loops enhances the robustness and traceability of automated vulnerability reproduction across diverse tasks and environments.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.