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

Perplexity has introduced a training method for its computer agent that incorporates data from both successful and failed user sessions. By using a technique called hint-guided self-distillation, the model learns from its previous errors to improve its reliability. In internal testing, this approach reduced the rate of tool-call failures for the agent from 2.24% to 1.77%.