STELLA: Towards a Biomedical World Model with Self-Evolving Multimodal Agents

摘要

ABSTRACT The staggering complexity of modern biomedical research has intensified the aspiration for a generalist “Biomedical World Model”, yet current AI agents remain constrained by static capabilities and a lack of self-evolution mechanisms. To bridge this gap, we present STELLA, a self-evolving multimodal agent designed to progressively refine its computational reasoning and physical execution through interaction. STELLA operates via a collaborative multi-agent framework (comprising Manager, Developer, Critic, Critic, and Tool Creation agents) that continuously updates reasoning templates and autonomously expands a dynamic “Tool Ocean”. We demonstrate STELLA’s capabilities on the created Tool Creation Benchmark, where it attains a score of 4.01/5 with 100% task completion, significantly outperforming state-of-the-art models including GPT-5, Claude 4 Opus, and Biomni. Beyond computational metrics, STELLA drives experimentally validated scientific discovery. In oncology, the agent identified Butyrophilin Subfamily 3 Member A1 (BTN3A1) as a novel negative regulator of natural killer (NK) cell function in acute myeloid leukemia (AML), verified via CRISPR knockout studies. In protein engineering, STELLA orchestrated a complete directed evolution workflow for the enzyme strictosidine synthase, identifying variants, notably M276L, exhibiting more than a two-fold improvement in catalytic activity. Finally, the system extends to physical laboratory automation by training Vision-Language-Action (VLA) models through a Decompose-Monitor-Recover mechanism, which increased success rates from 17% to 82%. By integrating autonomous tool evolution, biological discovery, and robotic control, STELLA offers a blueprint for a self-evolving world model in the life sciences.