← Back to feed

Perplexity published post-training research on April 22 that makes a direct and falsifiable claim: for search-augmented generation, the post-training pipeline matters more than which base model you select. Using open Qwen models as the base, Perplexity’s two-step SFT and RL process produces models that match or beat GPT-5.4 on some factual retrieval benchmarks at lower cost.

The Pipeline

Perplexity’s method runs in two stages.

Stage 1 — SFT: Trains the model to follow instructions accurately and stay calibrated. The primary goal is producing answers that are careful and appropriately scoped rather than verbose and approximate.

Stage 2 — RL: Trains specifically for search accuracy and tool-use efficiency. The design choices here are where the research gets specific.

Two problems Perplexity set out to solve in stage 2:

1. Search models call external tools too often, incurring latency and cost.
2. Search models treat verbosity as a proxy for quality — longer answers get rewarded even when they are wrong.

The solution is gated rewards: the style reward only activates when the answer is already factually correct. A model cannot earn points for being thorough while being wrong. Efficiency penalties are anchored to the best-performing trajectories within each training group, not applied as flat global penalties — which avoids punishing high-quality long answers alongside low-quality ones.

The Data Strategy

The training data follows the same two-objective logic:

• Verifiable multi-hop search questions for factual retrieval, where ground truth is checkable.
• Rubric-based conversational tasks to prevent the model from losing instruction-following ability when handling messy real-world traffic.

The combined dataset keeps both accuracy and chat quality from degrading against each other during RL — a standard failure mode in single-objective post-training.

What the Results Show

Post-trained Qwen models match or outperform GPT-5.4 on some factual tests, at lower operating cost. Perplexity does not publish absolute numbers in the summary, framing the result as benchmark-level parity rather than a comprehensive head-to-head.

The implication worth taking seriously: for search as a specific product capability, a company with access to a capable open-weight model and a well-designed RL recipe is not structurally dependent on frontier proprietary APIs. The base model is not the binding constraint.

What It Does Not Show

The research is internal Perplexity work tested on benchmarks of their choosing. Independent replication on search benchmarks they did not design has not been published. “Match or beat GPT-5.4 on some factual tests” is not the same as “equal GPT-5.4 on general factual retrieval.” The gap between internal benchmark performance and production search quality is well-documented across the industry.

That said, the gating mechanism — making style rewards conditional on factual correctness — is a specific and replicable technique. Teams running search-focused RL post-training on open models have a concrete design choice to evaluate.