Perplexity AI Implemented New Training Method

The company adopted a self-distillation training process to enhance tool-call accuracy.

Updated on Sept. 22, 2026 in Artificial Intelligence

Isometric editorial illustration of a modular data center server rack with dense cabling and illuminated indicators, representing technical AI infrastructure.
Perplexity AI has deployed a self-distillation training method designed to refine tool-call accuracy and reduce errors in its search performance. AI Illustration. Upload story photo >

Live Poll

Do you trust that AI search tools are becoming more reliable in providing accurate information?

Perplexity.AI has deployed a new self-distillation training method to improve its AI search performance. The company reported that this process successfully reduced tool-call failures by 21.2 percent in live testing environments.

Why it matters

High-quality tool use is a key differentiator for AI search products. Reducing unnecessary tool calls helps lower compute costs, minimize latency, and decrease error rates for end users.

The new method aligns with the DART-SD framework and utilizes the FRAMES benchmark for evaluating multi-source questions. The training pipeline integrates supervised fine-tuning with on-policy reinforcement learning using real-world user corrections.

The players

Perplexity.AI

Perplexity.AI is a technology company that develops an AI-powered search engine designed to provide direct, cited answers to user queries.

The details

The model undergoes supervised fine-tuning through curated examples of effective tool use before entering a reinforcement learning stage. This process incorporates actual tool error data and user feedback to refine how the AI handles complex, multi-source inquiries.

Timeline

  1. September 22, 2026: The company announced the implementation of the new training method.

The Big Picture

The development follows the DART-SD framework, which has become a key reference point for optimizing AI tool-use reliability.

Users can expect more accurate search results and fewer instances of the AI failing to retrieve data when answering complex questions. These technical improvements aim to streamline the search experience by reducing delays caused by unsuccessful model tool calls.

The takeaway

Reliable tool use remains a primary hurdle for developers striving to make generative AI search engines more dependable for daily information retrieval. Implementing reinforcement learning based on actual user corrections provides a scalable path for improving AI precision.

Further reading

Learn more about the latest industry developments in Artificial Intelligence.

Source note: This article includes information reported by Crypto Briefing.

Live Poll

Do you trust that AI search tools are becoming more reliable in providing accurate information?