Perplexity's just published a solid new research.

Claims that search quality comes mostly from post-training, not just from choosing a stronger base model.

The same open Qwen model can become much better at search after a special 2-step tune-up, so product quality is not fixed by the raw model alone.

That 2-step tune-up first teaches the model to follow instructions and stay careful, then teaches it to search more accurately and use fewer web lookups.

The key finding is that this recipe improved accuracy and efficiency at the same time, which is hard because search models often get better on 1 goal and worse on another.

The clever move is not just adding more reward terms, but gating them so style only counts when the answer is already correct.

That matters because search models have two expensive habits: they call tools too often, and they turn verbosity into a proxy for effort.

Perplexity tries to fix both with penalties anchored to the best working trajectories in each training group, not with blunt global punishments.

The data strategy follows the same logic.

Verifiable multi-hop search questions teach factual retrieval, while rubric-based chat tasks keep the model from forgetting how to follow instructions in messy real traffic.

The headline benchmark claims are strong, with post-trained Qwen models reportedly matching or beating GPT-5.4 on some factual tests at lower cost.