Event: Black Hat Asia
Location: Marina Bay Sands, Singapore (Thought Leadership Stage 1).
Format: Sponsored Session
Track: AI, ML & Data Science
Training LLMs with reinforcement learning (RL) has proven successful in many domains, yet there is limited public research on its application to computer security tasks. While any verifiable task can serve as a candidate for optimization, building effective training environments often presents significant challenges.
This presentation introduces methodologies to systematically design RL gyms tailored to security research. Practical guidance will be provided on designing robust reward functions along with strategies for developing efficient training environments. Further, detailed guidelines will be shared on effective RL model training, including considerations for distilling larger models and selecting key evaluation metrics beyond the reward function.
