Advancing search-augmented language models

Perplexity describes a two-stage post-training pipeline for open-source Qwen3-based web-search agents that separates deployment requirements from search optimization: supervised fine-tuning establishes instruction following, safety guardrails, language consistency, and production-style tool use, while on-policy reinforcement learning improves factual accuracy and tool efficiency. RL training combines synthetic multi-hop, web-verified question answering with rubric-based general-chat tasks that preserve performance on non-verifiable requests such as rewriting and planning, using a 90/10 mixture to prevent easier rubric tasks from dominating optimization. Its composite reward gates preference-based quality scores on verified correctness or complete rubric compliance and applies group-relative penalties for excessive tool calls and response length, aiming to limit reward hacking without discouraging needed exploration. Reported evaluations show improved search accuracy, abstention, language consistency, and user-preference metrics, with the post-trained Qwen model outperforming or matching cited competitors on several factual benchmarks while requiring fewer searches and lower estimated query costs. The authors note remaining challenges involving training–inference mismatch, multi-objective optimization at scale, and long-horizon workflows involving multiple tools.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.