Perplexity.AI Cuts Tool-Call Failures by 21% with New Training Method

Perplexity.AI has made significant strides in improving the accuracy of its AI models by reducing tool-call failures by 21.2% through a new self-distillation training method.

The approach, which aligns with the DART-SD framework, corrects failures without penalizing valid reasoning steps taken by the model.

Perplexity's implementation involves a two-stage training pipeline: supervised fine-tuning and on-policy reinforcement learning. The latter stage incorporates real-world user corrections and actual tool errors, allowing the model to learn from messy reality.

This is particularly interesting because it shows that the model isn't just training on idealized examples but is actually learning from real-world scenarios.