Perplexity cuts Computer's tool-call mistakes by training on real user sessions

Your move: A backend model change, nothing to set up

Perplexity Research described a new way the company trains the model behind Perplexity Computer. It learns from real user sessions, copying steps that worked and correcting specific mistakes, like a wrong tool call, even when the session succeeded overall. In a live test with about 100,000 users in each group, failed tool calls fell from 2.24% to 1.77% versus an earlier trained version, a drop of about 21%.

Why it matters

This is a backend improvement to how reliably Computer completes tasks, not a new feature or setting. If you use Computer for multi-step tasks, you should see fewer failed tool calls as this training rolls into the model you use. Perplexity has not released the method's code or weights, so there is nothing to install.

Source: Perplexity Research ↗

Published by Agent Log. The summary reports the linked source; “Why it matters” and the verdict are Agent Log's interpretation.

More Perplexity Computer updates →