CASCADE: Cumulative Agentic Skill Creation through Autonomous Development and Evolution

Xu Huang, Junwu Chen, Yuxing Fei, Zhuohan Li, Philippe Schwaller, Gerbrand CederarXiv 2025 · DOI · arXiv

Abstract

CASCADE is a self-evolving multi-agent framework designed to transition Large Language Models (LLMs) from static tool use to autonomous skill acquisition. By integrating continuous learning, self-reflection, and consolidated memory, CASCADE enables agents to master complex scientific software and workflows in real-time. Evaluated on SciSkillBench, a novel materials science benchmark, CASCADE achieved a 93.3% success rate, significantly outperforming native LLM baselines and demonstrating a path toward autonomous AI co-scientists.

Introduction

While Large Language Models (LLMs) are increasingly deployed as autonomous agents, they typically rely on a fixed set of predefined tools, such as calculators or specific search APIs. This creates a significant bottleneck in scientific research, where new software packages, undocumented databases, and complex experimental protocols emerge constantly. Current agents often struggle with these "out-of-distribution" tools, failing when faced with long-horizon tasks that require stitching together multiple libraries or debugging code for niche scientific packages.

To address this, we propose that a true AI "co-scientist" must move beyond mere tool use to skill acquisition—the ability to autonomously learn, codify, and reuse complex workflows. We introduce CASCADE (Cumulative Agentic Skill Creation through Autonomous Development and Evolution), a framework that allows agents to evolve their capabilities through experience.

The CASCADE Framework

CASCADE is a multi-agent system built around two core "meta-skills":

  1. Continuous Learning: The agent utilizes web search and code extraction to learn how to use new APIs and libraries in real-time.
  2. Self-Reflection: Rather than simple trial-and-error, the agent employs introspection, runtime probing, and knowledge graph exploration to diagnose failures and adapt its strategy.
Overview of the CASCADE framework: (a) The transition from tool use to skill acquisition; (b) The system architecture featuring the Orchestrator and Memory systems; (c) The DeepSolver multi-agent workflow.

Architecture and Workflow

The system is managed by an Orchestrator that directs tasks into two pathways:

  • SimpleSolver: A streamlined track for straightforward tasks or those the agent has already mastered.
  • DeepSolver: A sophisticated four-step workflow consisting of a Solution Researcher, Code Agent, Parallel Debug Agents, and an Output Processor. This pathway is designed for complex, novel problems requiring iterative refinement.

Crucially, CASCADE utilizes Consolidated Memory. When an agent successfully solves a problem or receives human guidance, it stores the executable "skill" in a database. This allows the system to solve similar problems faster in future sessions, creating a "compound effect" of knowledge.

Experimental Setup: SciSkillBench

To evaluate CASCADE, we developed SciSkillBench, a comprehensive benchmark of 116 tasks specifically tailored for materials science and chemistry. The benchmark covers six categories, ranging from data-oriented tasks (retrieval, analysis) to computation-oriented tasks (simulations using LAMMPS or ORCA). Tasks are categorized into two difficulty levels: Level 0 (procedural guidance provided) and Level 1 (high-level objectives only).

Results and Discussion

CASCADE demonstrated dramatic performance gains across various LLM backbones. Using GPT-5, CASCADE achieved a 93.3% success rate on SciSkillBench, nearly tripling the 35.4% success rate of the "Native" model.

Performance comparison of CASCADE (DeepSolver) against Native and Search&Debug baselines across multiple LLM backbones.

Key Findings

  • Efficiency and Scalability: CASCADE enabled smaller, more economical models (e.g., GPT-4.1-mini) to outperform much larger state-of-the-art coding models like Claude-Sonnet-4.5 on difficult tasks.
  • Parallel Debugging: By launching three different debug agents simultaneously—each using a different strategy like documentation review or runtime environment probing—CASCADE significantly increased the likelihood of resolving obscure errors in scientific code.
  • Real-World Application: Beyond benchmarks, CASCADE successfully controlled a physical autonomous laboratory (AlabOS) to synthesize inorganic powders, autonomously correcting errors in stoichiometric calculations during the process.

Conclusion and Future Work

CASCADE represents a significant step toward autonomous scientific discovery. By shifting the focus from tool use to skill acquisition, the framework allows AI agents to adapt to the rapidly evolving landscape of scientific software. Future work will focus on improving the agent's ability to generate truly novel scientific hypotheses and enhancing safety protocols for autonomous physical laboratory operations.

Summary updated

Abstract

Large language model (LLM) agents currently depend on predefined tools or early-stage tool generation, limiting their adaptability and scalability to complex scientific tasks. We introduce CASCADE, a self-evolving agentic framework representing an early instantiation of the transition from "LLM + tool use" to "LLM + skill acquisition". CASCADE enables agents to master complex external tools and codify knowledge through two meta-skills: continuous learning via web search, code extraction, and memory utilization; self-reflection via introspection, knowledge graph exploration, and others. We evaluate CASCADE on SciSkillBench, a benchmark of 116 materials science and chemistry research tasks. CASCADE achieves a 93.3% success rate using GPT-5, compared to 35.4% without evolution mechanisms. We further demonstrate real-world applications in computational analysis, autonomous laboratory experiments, and selective reproduction of published papers. Along with human-agent collaboration and memory consolidation, CASCADE accumulates executable skills that can be shared across agents and scientists, moving toward scalable AI-assisted scientific research.
BibTeX
@misc{huang2025cascade,
  author = {Xu Huang and Junwu Chen and Yuxing Fei and Zhuohan Li and Philippe Schwaller and Gerbrand Ceder},
  title = {{CASCADE: Cumulative Agentic Skill Creation through Autonomous Development and Evolution}},
  year = {2025},
  eprint = {2512.23880},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.23880},
}
Figure