Perplexity reports a 21% fall in failed tool calls after retraining its agent

single source· 1 articles · confidence: low · first seen 2026-09-25 14:30 UTC

What this means for you

Nothing to act on. This is an internal training run on one company's own agent: no published method, no weights, no date. The transferable point is that logged production failures are training data — if you serve an agent, you probably already have them.

Perplexity Research says retraining its Perplexity Computer agent cut failed tool calls from 2.24% to 1.77%, a 21.2% relative fall, in a live A/B test between two trained checkpoints. The method pairs rejection sampling fine-tuning, which keeps only generated attempts that pass a check, with hint-guided self-distillation — the model learns from its own outputs, with hints supplied, rather than from a separate teacher. Training draws on real user sessions, failures included. The write-up gives no evaluation date, no harness and no comparison against an untrained checkpoint.

Key facts

  • ·Perplexity Research trained a model inside Perplexity Computer on real user sessions, including failed ones source
  • ·Tool-call failures fell from 2.24% to 1.77% between two trained checkpoints in a live A/B test source
  • ·Perplexity reports the drop as a statistically significant 21.2% relative reduction source
  • ·The method pairs rejection sampling fine-tuning with hint-guided self-distillation source
  • ·The write-up gives no evaluation date, no harness and no comparison against an untrained checkpoint source

What the sources say

  • MarkTechPost — Short summary of Perplexity's post-training write-up, covering the method and the reported failure-rate drop

Sources

The original reporting. Follow these — they did the work.

← the wire