Credit Where It Matters: Dependency-Aware Policy Optimization for Terminal Agents
arXiv cs.AI · 3 Oct 2026, 03:24 AEST
Terminal-using agents benefit from reinforcement learning in multi-step tasks, but existing methods struggle with credit assignment. The proposed Dependency-Aware Group Policy Optimization improves learning by tracing command dependencies.
Key facts
- Terminal-using agents benefit from reinforcement learning (RL) in coding, debugging, and other multi-step terminal tasks.
- However, existing trajectory-level and step-level credit assignment methods do not explicitly trace the read-write dependencies through which commands affect the final outcome.
- In this paper, we propose Dependency-Aware Group Policy Optimization (DepGPO), which uses execution dependencies between commands to guide credit assignment for terminal agents.
- Extensive comparative experiments and ablation studies demonstrate that DepGPO improves task performance and training stability on complex terminal tasks.
Read aloud
Terminal-using agents benefit from reinforcement learning in multi-step tasks, but existing methods struggle with credit assignment. The proposed Dependency-Aware Group Policy Optimization improves learning by tracing command dependencies. Terminal-using agents benefit from reinforcement learning RL in coding, debugging, and other multi-step terminal tasks. However, existing trajectory-level and step-level credit assignment methods do not explicitly trace the read-write dependencies through which commands affect the final outcome. In this paper, we propose Dependency-Aware Group Policy Optimization DepGPO, which uses execution dependencies between commands to guide credit assignment for terminal agents. Extensive comparative experiments and ablation studies demonstrate that DepGPO improves task performance and training stability on complex terminal tasks.
Sources in this cluster
- Credit Where It Matters: Dependency-Aware Policy Optimization for Terminal Agents
arXiv cs.AI · 3 Oct 2026, 03:24 AEST