Perplexity reports fewer tool-call failures after hint-guided training
Perplexity Research has described a post-training method for its Computer agent that uses real user sessions, including unsuccessful ones. The approach combines rejection sampling fine-tuning with hint-guided self-distillation. In a live A/B test, tool-call failures declined from 2.24% to 1.77% between two trained checkpoints, a reported statistically significant relative reduction of 21.2%. Strong dissatisfaction changed from 2.58% to 2.54%, a difference that was not significant.
The method labels assistant turns for imitation, correction or context only. Correct turns in successful sessions receive cross-entropy training, while error turns with validated hints receive a forward Kullback-Leibler objective. The teacher sees the hint and the student does not; both process the recorded interaction with teacher forcing. Hints must rely on information available before the mistake, limiting hindsight bias.
In offline tests, the RFT-plus-OPSD checkpoint recorded a 0.87% tool-error rate, compared with 2.79% for stock GLM 5.2 and 1.35% for RFT alone, although the datasets differed. The training pipeline excludes opted-out users and sessions containing personally identifiable information. The post-trained weights and code have not been released; the model is available only as an option in Perplexity Computer, while GLM 5.2 is openly available on Hugging Face.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.