Perplexity Trains Its Computer Agent on Real Mistakes With Hint-Guided Self-Distillation
Perplexity reports a 21.2% live drop in Computer agent tool failures after hint-guided self-distillation.
Perplexity Research describes post-training its Computer agent on real GLM 5.2 user sessions, including failures, by pairing rejection-sampling fine-tuning with hint-guided on-policy self-distillation. The same checkpoint acts as a hinted teacher and an unhinted student; validated hints supply KL correction targets, while non-error turns in successful sessions receive cross-entropy loss. Offline tool-error rates were 2.79% for stock GLM 5.2, 1.35% for rejection sampling alone, and 0.87% for the combined method, though those checkpoints used different data. In a live test of about 100,000 users per condition, failures fell from 2.24% to 1.77%, a significant 21.2% relative reduction, while dissatisfaction did not change significantly, and weights were not released.
- Trains on real Computer sessions, including failed ones, using GLM 5.2.
- Hints become KL targets; successful clean turns are imitated with cross-entropy.
- Live tool-call failures fell from 2.24% to 1.77%, a 21.2% relative drop.
- Dissatisfaction was unchanged; weights and training code were not released.
- Vendors
- Perplexity
- Products
- Perplexity Computer
- Organizations
- Perplexity
- AI models
- GLM 5.2
From the original article
4 min read at marktechpost.com
Perplexity Research published a new post-training study. It trains a model inside Perplexity Computer on real user sessions, including failed ones. The method pairs rejection sampling fine-tuning with hint-guided self-distillation.
By Michal Sutter · marktechpost.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.