Perplexity improved its Computer model using real-world user sessions. It now learns from both successful and failed interactions. The method separates imitation of good steps from correction of mistakes. A later checkpoint cut tool-call failures by 21.2%. User corrections and tool errors provide the training signals. The pipeline excludes sessions with PII or from opted-out users. It uses rejection sampling fine-tuning plus hint-guided self-distillation. Successful sessions supply imitation and correction targets. Unsuccessful sessions supply only correction targets. The model acts as both teacher and student during training. The approach applies after reinforcement learning in synthetic environments.
What changed
Rejection sampling fine-tuning only imitated entire successful sessions without correcting intermediate mistakes.
What it unlocks
Training on both successful and unsuccessful real-world sessions to correct avoidable mistakes while imitating useful steps.
- 21.2% reduction in tool-call failures
What you need to act on it
- real-world user sessions with user corrections or tool errors
- sessions served by GLM 5.2
- privacy and opt-out filters applied
Sources