Details
- Perplexity says it post-trained a Computer model to learn from its own errors through hint-guided self-distillation.
- In a live A/B test, a later training checkpoint reduced tool-call failures by 21.2% relative to an earlier checkpoint.
- The training pipeline combines reinforcement learning in synthetic environments with real-world sessions that reveal failures outside designed scenarios.
- Sessions containing personally identifiable information and those from users who opted out of training are excluded from sampling.
- Unlike rejection-sampling fine-tuning, which mainly imitates successful sessions, the method uses successful actions and validated corrective hints from both successful and failed sessions.
- The annotation process links feedback to the decision responsible for an error and checks hints against information available before the mistake, limiting hindsight bias.
- GLM 5.2 evaluates the same recorded turn with and without a corrective hint; training then aligns the hint-free output with the improved behavior.
- With hints, the unchanged model avoided the original failure in 93.7% of cases, compared with 75.1% without the training effect.
- A separate live test showed tool-call failures falling from 2.24% to 1.77% between trained versions, without requiring hints during inference.
Impact
The result points to a practical way for agent developers to improve reliability without relying exclusively on larger models or costly inference-time supervision. Reducing failures from 2.24% to 1.77% could lower retries and operational overhead in tool-using systems, although the evidence comes from Perplexity’s internal tests rather than an independent benchmark. The privacy and opt-out filtering also show that production-session training is being paired with tighter data-governance controls.
