Autonomous scientific research at the heart of Perplexity Frontier Research

Perplexity’s Frontier Research program targets autonomous scientific research, with more than 10,000 GPUs and plans to publish its work.

Forming hypotheses, designing experiments, and revising conclusions: this research cycle is the focus of Perplexity Frontier Research. The new program aims to develop a system capable of producing interpretable scientific knowledge. In its research agenda, information retrieval becomes one component of a broader process that also includes challenging hypotheses and analyzing results.

Andrew Gordon Wilson leads the research effort. The team plans to publish papers, models, datasets, and benchmarks, and test its findings in Perplexity products. The announcement outlines a research and recruiting initiative, without setting a release date for an “automated scientist.” Agents that divide the research work

Collaboration between agents is central to the project. Some would explore a broad range of sources, while others would investigate a specific lead, propose experiments, or examine the conclusions reached. The team wants to study how to distribute these tasks and select models based on the desired quality and computational cost.

This coordination is intended to connect pretraining, reinforcement learning, and continual learning. The latter concerns the ability to learn from experience after deployment, retain useful information, and continue working through interruptions. The program also covers training and inference infrastructure, along with the security of environments where agents use tools. More than 10,000 GPUs to support the effort

The company says the research will be supported by more than 10,000 GPUs, without specifying the chip models or associated budget in its presentation. It is recruiting researchers and engineers in San Francisco, Palo Alto, New York, and London. The priority application deadline is December 5.

Public releases are expected to include scientific papers, model weights, and code repositories. The website already brings together work on agent memory, coordination between models, and execution environments. These references include research that predates the launch of Frontier Research. Measuring discovery rather than recall

Evaluation is a research area in its own right. The team wants to distinguish the ability to memorize an answer from the ability to reason, ask a relevant question, or propose an experiment. It also intends to measure whether a system provides interpretable scientific understanding beyond the accuracy of its predictions.

Among the highlighted projects, the WANDR benchmark, published in August, covers 500 information-gathering tasks. Agents must identify entities that meet specified criteria, investigate them through multiple searches, and return records backed by evidence. The test does not, by itself, measure autonomous scientific discovery, but examines one of its components: building a body of verifiable information. Its public evaluation framework refetches cited pages to check each record against its sources, then measures precision, coverage, and completeness.