Learning from Real-World Experience

Perplexity researchers describe a training pipeline that combines rejection sampling fine-tuning with hint-guided self-distillation to learn from both successful and unsuccessful real-world user sessions. By identifying avoidable mistakes through user complaints and tool errors and generating corrective hints for teacher-student training, the approach aims to improve agent decision-making. The company reports that post-training checkpoints reduced tool-call failures and estimated user dissatisfaction rates in live evaluations. Open article →