A good outcome does not prove a good process, and Perplexity Research has built a training method around that gap. The company published a post-training recipe that learns from real sessions of its computer agent, including the ones that failed, and reports a 21.2 percent relative cut in tool-call failures in a live A/B test.
Standard rejection sampling judges a session and copies the successful ones. Perplexity’s complaint is that an agent can botch a tool call, recover, and still finish correctly, so imitating the whole trajectory bakes in the error. Discarding failures wastes the clearest examples of what to avoid.
Its method splits the decision in two: which sessions deserve imitation, and which turns deserve correction. Good sessions supply both, bad ones only corrections. A hint is a short instruction built from information the model already had, naming the failed call and the validation error. Perplexity’s example is a search that passed a recency filter of year when the schema allowed only day, week or month.
Corrections use on-policy self-distillation. The same checkpoint replays the recorded turn twice, once with the hint and once without, and the teacher’s next-token probabilities serve as a soft target.
The weights stay private. Sessions containing personal data, and users who opted out, are excluded from the pipeline, which runs on GLM 5.2 inside Perplexity Computer.
