The error rate for tool calls in Perplexity's Computer agent fell from 2.24% to 1.77% during separate live tests after the company applied new post-training techniques to its GLM-5.2 model. The method uses rejection sampling fine-tuning and hint-guided self-distillation, also known as on-policy self-distillation, to teach the agent to learn from real user sessions by imitating successful trajectories and correcting mistakes such as bad tool calls.

Adding corrective hints allowed an unchanged model to avoid failures in 93.7% of cases, up from 75.1% without hints, while reducing overall failure rates by about 21% in live A/B tests. The training made agent trajectories more cost-efficient and showed gains in user satisfaction that are not yet statistically significant. Perplexity noted this approach provides a potential path for online and continual learning for AI agents.