Voices
Perplexity cuts live tool-call failures 21.2% by training its Computer agent on its own production traces
Perplexity's September 21 research post combines rejection-sampling fine-tuning on graded user sessions with hint-guided self-distillation. The same GLM 5.2 weights act as teacher (given a hint) and student (no hint), aligned through a KL loss. Hints alone raised tool-error avoidance from 75.1% to 93.7% on 985 held-out turns, and a live A/B test with about 100k users per arm showed a 21.2% relative drop in tool-call failures. Any agent team logging failed tool calls can reuse this recipe: bucket failed sessions by user correction and by tool error, then train on the fixed turns.
↳ Follow the thread