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

Perplexity Research improved its Perplexity Computer agent by training on real user sessions, including errors, using rejection sampling fine-tuning combined with hint-guided self-distillation. This approach reduced tool-call failures by 21.2% in live testing. The updated model runs only within Perplexity Computer and is not publicly released, though the base GLM 5.2 model is available on Hugging Face.