PACT: Privileged Trace Co-Training for Multi-Turn Tool-Use Agents
PACT improves multi-turn tool-use agents by using expert traces only as training signals, not for generating rollouts. It combines a trace-conditioned reinforcement learning surrogate with a component-aware supervised fine-tuning loss. This method keeps rollout generation prompt-only while guiding optimization through these two complementary signals, leading to state-of-the-art performance on benchmarks.