Keep Policy Gradient in Charge: Sibling-Guided Credit Distillation for Long-Horizon Tool-Use Agents

arXiv cs.CL · 18 hours ago

Researchers introduced Sibling-Guided Credit Distillation (SGCD), a method that improves reinforcement learning for agents performing multi-step tool-use tasks by using distillation to guide credit assignment rather than as a competing loss. The method improved performance on AppWorld from 42.9 to 45.6 on test_normal and from 24.7 to 27.0 on test_challenge, and on tau^3-airline from 0.583 to 0.602. This approach allows policy gradients to remain the primary actor update mechanism while distillation helps identify which actions lead to rewards.