1 paper · 1 filter
Shangao Li, Yao Zhang, Volker Tresp +1
LLM coding agents issue Bash commands through interfaces that may serialize, wrap, and reparse model output. Matched execution scores alone cannot distinguish command-generation er…