The Station: An Open-World Environment for AI-Driven Discovery

TL;DR

The Station introduces an open-world, autonomous multi-agent environment for scientific discovery that eliminates centralized orchestration and enables long, narrative-driven research. It demonstrates state-of-the-art performance across diverse benchmarks (circle packing, scRNA-seq batch integration, ZAPBench neural activity forecasting, Sokoban RL, and BEACON RNA modeling) and reveals emergent, cross-domain innovations such as density-aware batch integration and Fourier-based neural forecasting. The Open Station variant further exposes rich social dynamics, revealing how culture and metaphysical interpretations can emerge in the absence of external signals, posing both opportunities and safety challenges. Overall, the work argues that open-world environments can better harness the potential of advanced AI models for autonomous scientific contribution, by fostering exploration, collaboration, and emergent theory-building beyond rigid pipelines.

Abstract

We introduce the STATION, an open-world multi-agent environment for autonomous scientific discovery. The Station simulates a complete scientific ecosystem, where agents can engage in long scientific journeys that include reading papers from peers, formulating hypotheses, collaborating with peers, submitting experiments, and publishing results. Importantly, there is no centralized system coordinating their activities. Utilizing their long context, agents are free to choose their own actions and develop their own narratives within the Station. Experiments demonstrate that AI agents in the Station achieve new state-of-the-art performance on a wide range of benchmarks, spanning mathematics, computational biology, and machine learning, notably surpassing AlphaEvolve in circle packing. A rich tapestry of unscripted narratives emerges, such as agents collaborating and analyzing other works rather than pursuing myopic optimization. From these emergent narratives, novel methods arise organically, such as a new density-adaptive algorithm for scRNA-seq batch integration that borrows concepts from another domain. The Station marks a first step towards autonomous scientific discovery driven by emergent behavior in an open-world environment, representing a new paradigm that moves beyond rigid pipelines.

Figures (13)

• Figure 1: Illustration of the Station, an open-world multi-agent environment for autonomous scientific discovery. The Station is composed of multiple rooms, each serving a distinct purpose. Agents freely traverse between rooms and choose their own actions. Four example action paths are shown, such as agents performing independent research or collaborative analysis. These paths are unscripted, and actual trajectories are often much more complex and span hundreds of steps.
• Figure 2: Progress curve on the Circle Packing task.
• Figure 3: Illustration of the density-adaptive, batch-aware algorithm discovered in the Station: dense regions mix across batches, sparse regions connect within batches.
• Figure 4: Progress curve on the batch integration task.
• Figure 5: Performance comparison on the batch integration task. The "Perfect embedding by celltype with jitter" method serves as a positive control, representing the best possible performance. Conversely, "Shuffle integration by batch" serves as a negative control. The "Overall score" is the average of all metrics across all datasets. Each "Datasets" column displays the mean of all metrics for that specific dataset, while each "Metrics" column displays the mean of that specific metric across all datasets.