§ feed · storyline
Learning from Real-World Mistakes
Perplexity Research post-trained its Computer model on user corrections and tool failures, cutting tool-call errors by 21.2% from 2.24% to 1.77%, though user dissatisfaction gains were not confirmed.
Perplexity Research described post-training a Perplexity Computer model using user corrections and tool failures. The later checkpoint reduced tool-call errors from 2.24% to 1.77%, a statistically significant relative reduction of 21.2%, though the online tests did not establish lower user dissatisfaction.
§ sources1 publication · timeline below
- perplexity.aiLearning from Real-World Mistakesprimary