The authors developed LENS, a reinforcement learning (RL)-based framework that treats adversarial input generation as an RL problem.
Black-box RL method finds large performance gaps in LLM network control
LENS uses reinforcement learning to discover adversarial inputs that cause LLM-based traffic engineering and GPU allocation systems to fail significantly, even when those systems report high average accuracy.
Big Tech
Mahdi Alizadeh · Pooria Namyar · Arvin Ghavidel · Ramesh Govindan
University of Southern California · Microsoft Research
Research Digest··3 min read
Thread:Agent Security & Attacks
Alizadeh et al.
§