Tag
The paper introduces BernBRS, a scalable framework for certified backward reachability analysis of DeepBernNets, computing certified inner and outer approximations of the Backward Reachable Set via activation-pattern-based refinement of the input space cover, demonstrated on Half-Cheetah and TinyBERT-derived models.
This paper learns the constrained region of attraction for transverse feedback linearization under controller singularities by decomposing the Zubov value function into an analytic quadratic cost-to-go and a learned safety value function via fitted value iteration, achieving substantially better coverage than quadratic and PINN baselines on unicycle and quadrotor path-following tasks.
SCOUT introduces a two-stage agentic safety verifier that combines reasoning-intensive rubric generation with tool-intensive evidence gathering to detect unsafe behaviors in computer-use agents, improving unsafe detection on AutoElicit-Bench and reducing unsafe execution rates from 30.2% to 17.2% via test-time reflection.
PolicyGuide compiles domain policies into workflow graphs and uses a proactive verifier to guide LLM agents through multi-step procedures, improving policy compliance across various benchmarks.