Table of Contents
Perplexity has launched Perplexity Frontier Research, an initiative to develop AI systems that can carry out scientific research autonomously, from identifying unanswered questions to testing hypotheses and refining their conclusions.
The effort brings together work on model architectures, learning algorithms, multi-agent collaboration, and evaluations, with support from more than 10,000 GPUs. Perplexity plans to publish papers, models, data, code, and benchmarks, and test the resulting advances in its products.
Andrew Gordon Wilson is leading the frontier research effort. He joined Perplexity as research lead in August and said he wanted to pursue continual learning and agent collaboration.
The objective extends beyond finding and summarizing existing information. Perplexity wants to automate a research cycle in which a system asks a useful question, develops a possible explanation, designs an experiment, examines the evidence, and updates its understanding. The intended output is new, interpretable knowledge: explanations and testable ideas, rather than predictions alone.
Coordinating Models Around a Research Question
Perplexity’s research agenda groups the work into six areas: multi-agent orchestration, pre-training, continual learning, reinforcement learning, systems, and frontier evaluations. Orchestration is intended to connect these efforts into a system whose agents perform complementary parts of an investigation.
Some agents would search broadly across thousands of sources, while others would investigate a particular thread in depth. Additional agents could propose experiments, critique reasoning, or develop hypotheses. The research challenge includes deciding when collaboration improves an answer, which model should handle a task, and when further searching is worth the computational cost.
One proposed approach is advisor escalation: a small model running on a user’s device asks a larger model for help only when that assistance is useful. Rather than routing every task through the most expensive model, the system would allocate resources according to the problem.
Perplexity’s target is the best achievable quality at a user’s chosen cost. In technical terms, it wants output on the quality-cost Pareto frontier, where improving quality would require spending more, or reducing cost would sacrifice quality.
The team also plans to study how much of an agent’s ability should reside in its learned model weights and how much should come from the surrounding software, often called the harness. That software manages tools, context, and the sequence of actions a model takes. Other questions include whether specialized models can coordinate more efficiently and whether models should be trained jointly with their eventual orchestration roles in mind.
Training and Evaluating Research Models
The pre-training agenda focuses on efficient, specialized open-weight models, new architectures, synthetic data, and how results from small experiments transfer to larger training runs. Perplexity wants to investigate whether generated training data can be more useful than real data and whether changes to training methods can alter how performance improves with scale.
Continual learning addresses what happens after deployment. The planned systems would seek information they do not already know, retain useful experience, and carry work across interruptions. For a research agent, that could mean accumulating progress over days rather than restarting an investigation at each new session.
Reinforcement learning, in which models learn from feedback on their actions, introduces another set of problems. An investigation might involve thousands of actions before producing a useful result. The team wants to understand which actions deserve credit when the feedback arrives only at the end, how well simulated training environments transfer to the live web, and how to design rewards that resist exploitation.
Measuring success is a research area of its own. Perplexity plans evaluations that distinguish reasoning from memorization and scientific understanding from predictive accuracy. Its agenda uses the example of a model that can predict a particle’s trajectory: producing accurate positions is different from deriving equations of motion or a theory that explains them.
The evaluation goal is to test whether systems can ask productive questions, propose experiments, improve their own abilities, and generate explanations that teach people something new. That connects the measurement work directly to the proposed scientist’s tasks, rather than treating a higher score on an existing test as the endpoint.
Open Research, Computing Infrastructure, and Hiring
The systems work covers training and inference, communication between GPUs, and secure environments for long-running agents. Fewer computational operations do not automatically make a model faster if it spends most of its time moving data between devices.
Infrastructure for autonomous research also needs to preserve an agent’s work and support extended tool use. The agenda includes isolated execution environments that let agents continue working while keeping credentials and the host system outside their control.
Perplexity says findings will feed into products used by millions of people. Alongside its planned releases, the research site links existing work on retrieval evaluations, model orchestration, embeddings, GPU communication, and agent runtimes, which give researchers concrete examples of the technical areas the initiative covers.
The company is hiring researchers and engineers in San Francisco, Palo Alto, New York, and London. Its Frontier Research application has a December 5 priority deadline.
A separate Research Fellowship offers predoctoral and postdoctoral tracks, with a three-month full-time commitment and opportunities to extend or convert to a permanent research role. That program accepts applications on a rolling basis and is based in Perplexity’s San Francisco, Palo Alto, and New York offices.