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

Digest AI · Agents & Tools · published 2026-09-25T14:30:26Z

Canonical: https://digestai.news/story/perplexity-trains-agent-on-real-mistakes-with-hint-guided-self-distill

## Summary

Perplexity Research published a study on a new post-training method for its agent inside Perplexity Computer. The approach uses rejection sampling fine-tuning combined with hint-guided self-distillation to improve tool-call accuracy. In a live A/B test, tool-call failures dropped from **2.24%** to **1.77%** between two trained checkpoints, marking a **21.2% relative reduction** the company calls statistically significant.

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.

## Key points

- Perplexity’s agent training method reduces tool-call failures from 2.24% to 1.77% in live tests, a 21.2% relative drop
- Hints—short corrective instructions—guide self-distillation by comparing teacher (with hint) and student (without) model outputs
- Base model GLM 5.2 remains open on Hugging Face, but post-trained weights and code are not released

## Why it matters

This method could improve agent reliability by learning from failures, not just successes, reducing avoidable errors in real-world tool interactions.

## Sources

1. [Perplexity Trains Its Computer Agent on Real Mistakes With Hint-Guided Self-Distillation](https://marktechpost.com/2026/09/25/perplexity-trains-its-computer-agent-on-real-mistakes-with-hint-guided-self-distillation) (MarkTechPost, 2026-09-25)

## Cite

Digest AI, "Perplexity trains agent on real mistakes with Hint-Guided Self-Distillation", 25 September 2026, https://digestai.news/story/perplexity-trains-agent-on-real-mistakes-with-hint-guided-self-distill

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/perplexity-trains-agent-on-real-mistakes-with-hint-guided-self-distill.json