CoTrace: Data Recipes for Training Terminal Agents with Harness-Model Co-Evolution

In 30 Sekunden

Was passiert ist

CoTrace method improves terminal agents by matching training trajectories to the runtime harness that will execute them, advancing Qwen3.5-9B from 78 to 90 solved tasks.

Warum es zählt

Matters for engineers building autonomous agents that interact with terminals or APIs, where the execution environment affects model performance and generalization.

Achtung

The approach requires careful alignment between training and evaluation harnesses; mismatched runtimes cause procedural breakdowns, limiting out-of-distribution transfer capabilities.