JET: Just-Enough Exploration Tuning for Efficient Tool-Calling Agents in Repository-Level Question Answering
JET cut tool calls by 11–33% in repository QA tests while largely preserving answer scores.
The preprint finds that agents keep exploring code after collecting enough information to answer, and supervised fine-tuning can pass that behavior on. JET cuts training trajectories at the point where sufficient information is available, adds a rationale for answering there, and gives that rationale extra weight during training. Tests covered two benchmarks and two open-weight 8B models. With expert trajectories, JET used 18–28% fewer tool calls than standard supervised fine-tuning while retaining most of its score gains; inference costs fell in both settings.
ArXiv · AI/CL/LG's note
The preprint finds that agents keep exploring code after collecting enough information to answer, and supervised fine-tuning can pass that behavior on. JET cuts training trajectories at the point where sufficient information is available, adds a rationale for answering there, and gives that rationale extra weight during training. Tests covered two benchmarks and two open-weight 8B models. With expert trajectories, JET used 18–28% fewer tool calls than standard supervised fine-tuning while retaining most of its score gains; inference costs fell in both settings.
ArXiv · AI/CL/LG's note
score 5