From the source
Lead story
Ask your AI
Top stories
Models & availability
Latest
Perplexity· PerplexityLearning from Real-World Mistakes
Training method reduces tool-call failures by 21.2% using real-world user corrections and errors.
Perplexity describes a training method combining rejection sampling fine-tuning with hint-guided self-distillation to learn from real-world user sessions.
The approach distinguishes between successful and unsuccessful sessions, using user corrections and tool errors as training signals to reduce tool-call failures by 21.2% relative to an earlier checkpoint.
From the source
In live use, the later trained checkpoint reduced tool-call failures by 21.2% relative to an earlier trained checkpoint.
