Perplexity Learns From Failures

🔥 What happened Perplexity published a post-training method that trains agents on real user sessions, including failed ones. Instead of imitating only successful trajectories, it corrects mistakes using validated hints. In a live A/B test, tool-call failures dropped from 2.24% to 1.77%. 💡 Why it matters That's a 21.2% relative reduction — statistically significant, but the weights and training code stay closed. The base checkpoint, GLM 5.2, is open; the method isn't. For agent builders, the technique is compelling but not reproducible. ⚡ Our take Perplexity shows how to learn from failure — then hides the recipe. Open-source communities will rebuild this, whether Perplexity likes it or not.