When an off-the-shelf coding agent is used directly as a robot policy, observing a browser-based 3D interface through screenshots and acting by posing a virtual target gripper through a few tools, the agent's harness, its prompts, tools, and control rules, largely determines success, and until now it has been written b...
Zhi-Jie Wei, Ferris Tan, Jing-Hui Wang et al.· 2 citations
This work presents Benchproofer, a pipeline that turns a coding task with a known correct patch into a formally verified one, which writes a specification for the new code, summarizes the existing functions that code calls with axioms, and admits an instance only after mechanical and adversarial gates agree.
George Ma, Benjamin Mikek, Hao-Yu Li et al.· 2 citations
Evo-Bench is the first benchmark designed to evaluate models'intrinsic harness-evolving capabilities across Search, Office, and General agent domains, and exposes critical temporal anomalies like early saturation, while demonstrating that the synthesized harnesses act as highly transferable reasoning structures, consis...
