Perplexity.AI has implemented a new self-distillation training method that reduced tool-call failures by 21.2% during live testing. The approach is based on the DART-SD framework, which is designed for multi-turn tool-calling agents and aims to correct errors without affecting valid reasoning steps.
The training pipeline consists of two stages. Initially, the model undergoes supervised fine-tuning using curated examples of correct tool usage. This is followed by on-policy reinforcement learning, where the model is trained using real-world user corrections and actual tool errors, including timeouts and formatting issues.
The company is focusing on improving accuracy while reducing the number of tool calls to optimize compute usage and latency. This strategy is also reflected in their work on the FRAMES benchmark, which evaluates the performance of AI systems in handling complex multi-source queries under controlled tool budgets.
