Perplexity Trains Its Computer Agent on Real Mistakes With Hint-Guided Self-Distillation

A computer agent was trained using a method that learns from both successful and failed real user sessions, rather than only successful ones. The training approach pairs rejection sampling fine-tuning with hint-guided self-distillation, where hints are corrective instructions based on information the model already had access to, and the same model checkpoint serves as both teacher and student during training. In live testing, tool-call failures decreased from 2.24% to 1.77% between two trained checkpoints, representing a statistically significant 21.2% relative reduction. The trained model runs only within a specific product and has not been released publicly, though the underlying base model is available openly.