Perplexity Trains Its Computer Agent on Real Mistakes With Hint-Guided Self-Distillation
21.2% fewer tool-call failures. Perplexity's Computer agent learns from real mistakes—including failures—using hint-guided self-distillation.
Why it matters
Perplexity deployed a production technique to improve agent reliability by training on real user sessions and failed interactions. This is a concrete agent reliability engineering story: measured tool-call failure reduction (2.24% → 1.77%) in live A/B test, with a reproducible post-training method combining rejection sampling and self-distillation.
The key facts
5 to knowTool-call failure rate: 2.24% → 1.77% (21.2% reduction claimed as statistically significant)
Method: rejection sampling fine-tuning paired with hint-guided self-distillation
Training data: real Perplexity Computer user sessions, including failed interactions
Deployment: live A/B test between two model checkpoints
Source: Perplexity Research post-training study; published via MarkTechPost
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Perplexity Research published a new post-training study. It trains a model inside Perplexity Computer on real user sessions, including failed ones. The method pairs rejection sampling fine-tuning with hint-guided self-distillation. In a live A/B test, tool-call failures fell from 2.24% to 1.77%…