S1-NexusAgent: a Self-Evolving Agent Framework for Multidisciplinary Scientific Research

TL;DR

S1-NexusAgent introduces a self-evolving, cross-domain scientific agent built on a Plan-and-CodeAct architecture with an inner–outer dual-loop that separates global planning from subtask execution. It features an extensible MCP-enabled tool ecosystem with intent-aware dynamic hot-plugging and a sparse context management framework to sustain long-horizon reasoning across thousands of tools. A trajectory-evaluation-based self-evolution (TE-SE) module continuously refines behavior by distilling high-quality execution trajectories into reusable Scientific Skills, enabling cross-domain knowledge transfer. Empirical results on Biomni-Eval, ChemBench, and MatSciBench demonstrate state-of-the-art performance and robust generalization, validating the framework's practicality for real-world multidisciplinary research.

Abstract

Modern scientific research relies on large-scale data, complex workflows, and specialized tools, which existing LLMs and tool-based agents struggle to handle due to limitations in long-horizon planning, robust goal maintenance, and continual learning from execution. To address these issues, in this work, we propose S1-NexusAgent, a self-evolving agent framework designed for multidisciplinary scientific research. S1-NexusAgent adopts a hierarchical Plan-and-CodeAct execution paradigm, decoupling global scientific planning from subtask-level tool execution through a dual-loop architecture, thereby enabling stable modeling of complex research workflows. The system natively supports the Model Context Protocol (MCP), integrates up to thousands of cross-disciplinary scientific tools, and achieves efficient orchestration of heterogeneous research tools via intention-aware dynamic tool retrieval and hot-plug mechanisms. To address long-context and large-scale data challenges in scientific settings, S1-NexusAgent introduces object-reference-based sparse context management, which enables sub-task context isolation and intermediate result compression. Building on this, a Critic Agent automatically evaluates complete execution trajectories and distills high-quality research paths into reusable Scientific Skills, forming a closed loop for continuous self-evolution, which is valuable for sustainable and long-horizon scientific research. Experiments on authoritative scientific benchmarks involving long-horizon planning and complex specialized tool orchestration, including biomini-eval (biology), ChemBench (chemistry), and MatSciBench (material science), demonstrate that S1-NexusAgent achieves state-of-the-art performance, validating its effectiveness and generalization capability in complex scientific tasks.

Figures (7)

• Figure 1: Architectural Overview of S1-NexusAgent.
• Figure 2: Overview of the Intent-aware Dynamic Hot-Plugging (DHP) mechanism for scalable tool retrieval. The framework first performs domain-level filtering based on structured intent to prune irrelevant tool categories. Subsequently, fine-grained semantic matching is conducted within candidate domains to facilitate the plug-and-play injection of the most relevant tools into the execution context. This on-demand strategy effectively mitigates contextual noise and prevents reasoning interference, ensuring robust tool invocation in large-scale scientific toolsets.
• Figure 3: The modular and extensible scientific tool ecosystem of S1-NexusAgent. Scientific capabilities are abstracted into three functional layers: domain-specific tools (e.g., bioinformatics, materials simulation), literature retrieval services, and general utilities. By adopting CodeAct as the unified action paradigm and maintaining compatibility with standardized tool protocols, S1-NexusAgent provides a secure, sandboxed environment for programmatic tool interaction, enabling seamless integration of diverse scientific instruments and databases without model retraining.
• Figure 4: The Inner--Outer Dual-Loop architecture of S1-NexusAgent. The Outer Loop, governed by the Planner Agent, maintains global research objectives and regulates high-level strategy through recursive task decomposition. Complementing this, the Inner Loop leverages the CodeAct paradigm for localized exploration, enabling high-frequency trial-and-error, self-correction, and parameter tuning within specific subtasks. The synergy between these loops allows the system to adaptively scale execution depth and planning granularity according to task complexity.
• Figure 5: The Context Management framework for long-horizon scientific discovery. To prevent context explosion and reasoning drift, the system integrates four specialized mechanisms: (1) Object-Level Referencing via lazy loading of large data artifacts; (2) Subtask-Level Isolation to minimize inter-stage interference; (3) Execution Trace Compression through structured distillation of historical steps; and (4) Planner-Aware Augmentation that feeds high-value signals back to the global state. Together, these mechanisms ensure system scalability and reasoning coherence across long-duration research workflows.