PACT: Privileged Trace Co-Training for Multi-Turn Tool-Use Agents

A new training framework called PACT aims to improve how AI agents handle multi-step tasks that require tool use, according to a paper submitted to arXiv in 2026 [1]. The method uses expert demonstrations as optimization signals during training without relying on them when the agent is actually running [2]. Multi-turn tool-use agents must reason, call external tools, and adapt to observations across several interaction turns [2]. Training these agents has presented a persistent trade-off. Reinforcement learning (RL) matches the prompt-only inference setting but often suffers from sparse rewards and weak credit assignment, while supervised fine-tuning (SFT) on expert traces provides dense process supervision but can over-constrain the model to fixed trajectories [2]. The PACT framework, short for Privileged trAce Co-Training, attempts to resolve this by keeping rollout generation prompt-only and using expert traces solely as training-time optimization signals [2]. It does so through two complementary signals: a trace-conditioned RL surrogate that evaluates prompt-only rollouts under expert-trace context, and a component-aware SFT loss that supervises reasoning prefixes and tool-calls with annealed strength [2]. To reduce over-reliance on the training-only trace context, PACT introduces a prompt-only anchoring mechanism [2]. The authors also provide a latent-trace view that connects the two trace-based objectives and explains how expert traces can guide optimization without being used during rollout generation [2]. Experiments were conducted on three benchmarks — FTRL, BFCL, and ToolHop — and the results showed that PACT consistently improves over strong SFT- and RL-based baselines [2]. The paper was posted on arXiv, an open-access repository of electronic preprints that has hosted scientific papers since 1991 and now receives about 24,000 submissions per month [8]. arXiv papers are moderated but not peer reviewed [8]. The repository supports community-developed tools through its arXivLabs framework, which allows collaborators to build features such as bibliographic explorers and code finders directly on article pages [6][7].