Perplexity trains agent on real mistakes with Hint-Guided Self-Distillation

The method distinguishes between successful and failed sessions, applying cross-entropy loss to correct turns in successful sessions and KL divergence loss to error turns with validated hints. Hints are short corrective instructions grounded in the model’s existing knowledge. Perplexity notes the base model, GLM 5.2, remains openly available on Hugging Face, but the post-trained weights and training code are not released. Offline tool-error rates fell from 2.79% for stock GLM 5.2 to 1.35% with rejection sampling fine-tuning alone, and further to 0.87% with the combined approach. However, live user dissatisfaction rates showed no significant change, remaining around 2.58% to 2.54%.

The study excludes sessions with personally identifiable information and uses LLM judges to validate task difficulty and root-cause analysis of complaints. Perplexity emphasizes that hints improve model behavior before training, with the base model avoiding original failures in 93.7% of held-out tool-error turns when provided with hints.