TLDR Tech / AI
Advancing Search-Augmented Language Models
Perplexity's two-stage pipeline for search-augmented language models enhances factual accuracy, user preference, and tool-use efficiency using initial Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL). This method, starting with Qwen3 models, separates compliance from search improvement, achieving high accuracy without compromising guardrails. The models demonstrated improved accuracy on benchmarks like FRAMES and FACTS OPEN, with reduced cost per query and better tool usage efficiency compared to models like GPT-5.4.
Read the full story