DCAS: Decoupling CLI Agent Scaffolding to Internalize Planning across Scaffolds
Fine-tuning under OpenHands appears to teach scaffold-specific planning habits that fail elsewhere.
The paper introduces DCAS, an interception layer that lets researchers pair CLI agent scaffolds with backend models without changing the scaffold. Its experiments find that models fine-tuned on OpenHands trajectories do well there but drop under other scaffolds, while base models do not show the same split. A planning-source intervention points to planning quality as a major driver of the gap. A small set of DCAS-collected, planning-aware trajectories improved performance across non-training scaffolds. HF Daily Papers' note
The paper introduces DCAS, an interception layer that lets researchers pair CLI agent scaffolds with backend models without changing the scaffold. Its experiments find that models fine-tuned on OpenHands trajectories do well there but drop under other scaffolds, while base models do not show the same split. A planning-source intervention points to planning quality as a major driver of the gap. A small set of DCAS-collected, planning-aware trajectories improved performance across non-training scaffolds. HF Daily Papers' note
score 5