CASCADE: Cumulative Agentic Skill Creation through Autonomous Development and Evolution

TL;DR

The paper addresses the limitations of current LLM agents that rely on predefined tools by introducing CASCADE, a self-evolving framework for cumulative skill creation. CASCADE leverages two meta-skills—continuous learning via web search and code extraction, and self-reflection via introspection and knowledge-graph exploration—to acquire executable skills and master external tools, while maintaining memory for persistent learning. A new benchmark, SciSkillBench, evaluates Bio/Chem/materials science tasks across data and computation domains and shows DeepSolver achieving 93.3% success with GPT-5 versus 35.4% without evolution, including autonomous lab demonstrations and reproduction of published results. The work demonstrates that executable, shareable skills can be accumulated and transferred across agents and scientists, moving toward scalable AI-assisted scientific research with domain-agnostic tool integration and end-to-end problem solving.

Abstract

Large language model (LLM) agents currently depend on predefined tools or brittle tool generation, constraining their capability and adaptability 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 and code extraction, and self-reflection via introspection and knowledge graph exploration, among 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.

Figures (5)

• Figure 1: The "LLM + skill acquisition" paradigm and the CASCADE architecture.a, A puzzle-solving metaphor of the "LLM + tool use" versus the "LLM + skill acquisition" paradigm. On the left, agents rely on human experts to curate external tools. On the right, CASCADE showcases its ability to adeptly craft customized tools from complex external components, facilitating use by both human experts and LLM agents. b, The architecture of CASCADE. CASCADE facilitates multi-turn dialogues with human scientists through an interactive web interface, with persistent session management and consolidated memory in vector and graph databases. The Orchestrator agent within CASCADE selects between two solution pathways, the SimpleSolver or the DeepSolver, based on adaptable memory and task difficulty. c, DeepSolver architecture. DeepSolver coordinates four specialized agents that collaboratively solve complex tasks while autonomously acquiring new tools and skills. It follows a sequential workflow: the Solution Researcher generates the initial code solution; the Code Agent executes the code; if debugging is required, the Debug Agents intervene; and finally, the Output Processor Agent processes the results.
• Figure 2: Task diversity in SciSkillBench and performance analysis.a, Overview of the diverse tasks involved in SciSkillBench, comprising two main types: data-oriented tasks (76 in total) and computation-oriented tasks (40 in total). There are six specific categories with different proportions, namely: data retrieval (19.0%), data analysis (20.7%), data management (3.4%), data processing (22.4%), simulation (27.6%), and specialized models and toolkits (6.9%). The associated major databases, packages, and software for each category are listed, alongside examples of quantities that the benchmarked system is required to obtain. b, Pass@3 accuracy against task difficulty for DeepSolver, Search&Debug Baseline (S&D), Native, and Claude Code Baseline. Each backbone model is represented by a distinct color and marker shape. Claude Code Baseline (red star) is shown in the DeepSolver section.
• Figure 3: Piezoelectricity determination and prediction of machine learning interatomic potential systematic errors.a, Determining whether a structure exhibits piezoelectricity. Given this task, CASCADE executed the necessary code and reached the correct conclusion. b, Hypothesis formulation, experimental execution, and data analysis of systematic differences in density and volume per atom predictions using machine learning interatomic potentials trained on different density functional theory (DFT) functional data. CASCADE not only provided a reasonable and accurate solution but also generated compelling visualizations, with the four plots from left to right being: scatter_volume_per_atom.png, scatter_density.png, hist_delta_volume_per_atom.png, and hist_delta_density.png.
• Figure 4: CASCADE integrated into the autonomous lab materials discovery loop. CASCADE submitted the synthesis task for Li$_2$Fe$_{0.8}$Ni$_{0.2}$Cl$_4$ to the autonomous lab (A-Lab) platform, allowing us to carry out the compound's synthesis, characterization, and electrochemical impedance spectroscopy (EIS) measurements. The acquired experimental data were fitted using CASCADE to determine the ionic conductivity. The bottom right corner shows a Nyquist plot visualizing the quality of the data fitting performed by CASCADE.
• Figure 5: Human-agent collaboration with memory capability for reproducing published content.a, Identification of issues in user queries. When asked to generate text descriptions with a mismatched material and Materials Project ID, CASCADE directly identified the problem and requested clarification. b, Calculation of average voltages. CASCADE computed the average voltages over specified ranges, $0 < x < 1$ and $0.5 < x < 1$, for Li$_x$CoO$_2$ during multi-turn interactions with users. c, Reproducing published work. CASCADE successfully calculated the average voltages over specified ranges for Li$_x$CoO$_2$, using four SevenNet models. Then, it generated plots similar to those in the referenced article. The bottom left plot is labeled average_voltage_models_labeled.png, and the bottom right plot is average_half_voltage.png.