EvoMaster: A Foundational Agent Framework for Building Evolving Autonomous Scientific Agents at Scale
| Authors | Xinyu Zhu et al. |
| Year | 2026 |
| HF Upvotes | 3 |
| arXiv | 2604.17406 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
The convergence of large language models and agents is catalyzing a new era of scientific discovery: Agentic Science. While the scientific method is inherently iterative, existing agent frameworks are predominantly static, narrowly scoped, and lack the capacity to learn from trial and error. To bridge this gap, we present EvoMaster, a foundational evolving agent framework engineered specifically for Agentic Science at Scale. Driven by the core principle of continuous self-evolution, EvoMaster empowers agents to iteratively refine hypotheses, self-critique, and progressively accumulate knowledge across experimental cycles, faithfully mirroring human scientific inquiry. Crucially, as a domain-agnostic base harness, EvoMaster is exceptionally easy to scale up -- enabling developers to build and deploy highly capable, self-evolving scientific agents for arbitrary disciplines in approximately 100 lines of code. Built upon EvoMaster, we incubated the SciMaster ecosystem across domains such as machine learning, physics, and general science. Evaluations on four authoritative benchmarks (Humanity's Last Exam, MLE-Bench Lite, BrowseComp, and FrontierScience) demonstrate that EvoMaster achieves state-of-the-art scores of 41.1%, 75.8%, 73.3%, and 53.3%, respectively. It comprehensively outperforms the general-purpose baseline OpenClaw with relative improvements ranging from +159% to +316%, robustly validating its efficacy and generality as the premier foundational framework for the next generation of autonomous scientific discovery. EvoMaster is available at https://github.com/sjtu-sai-agents/EvoMaster.
Engineering Breakdown
Plain English
EvoMaster is a new agent framework designed to enable large language models to conduct scientific research autonomously through iterative hypothesis refinement and self-critique. The paper addresses a critical gap: existing agent frameworks are static and don't learn from experimental failures the way human scientists do. The authors built EvoMaster as a domain-agnostic platform that allows agents to accumulate knowledge across multiple experimental cycles, progressively improving their approach through trial-and-error feedback loops. This enables scaling agentic science workflows across diverse research domains without domain-specific engineering.
Core Technical Contribution
The core innovation is an evolving agent architecture that implements continuous self-evolution as a first-class design principle, rather than treating agents as fixed inference pipelines. Unlike prior agent frameworks that execute a static plan, EvoMaster maintains internal state across experimental cycles and uses self-critique mechanisms to update hypotheses based on outcomes. The framework decouples the scientific hypothesis-refinement loop from the underlying LLM, making it composable and domain-agnostic—developers can apply the same EvoMaster harness to chemistry, biology, physics, or other scientific domains. This is architecturally different from prompt-chaining or multi-step reasoning approaches because it explicitly models the iterative scientific method with knowledge accumulation.
How It Works
EvoMaster operates as a state-machine controller wrapping an LLM, cycling through four key phases: hypothesis generation, experiment execution, result observation, and self-critique refinement. In the hypothesis phase, the agent proposes a testable scientific claim using the LLM. The experiment phase executes this claim in a domain-specific simulator or real environment, capturing structured results. The observation phase parses outcomes and detects success or failure conditions. Crucially, the self-critique phase uses the LLM to generate explanations for why the experiment succeeded or failed, and these critiques are stored in a persistent knowledge buffer that informs the next hypothesis cycle. The framework maintains a growing context window of past experiments, failures, and lessons learned—enabling the agent to recognize patterns and avoid repeated mistakes across multiple rounds.
Production Impact
For teams building scientific discovery pipelines, EvoMaster eliminates the need to hand-craft multi-stage reasoning prompts for each new domain—you define the experiment execution interface and the framework handles the iterative refinement loop. This reduces development time from weeks of prompt engineering to days of simulator/API integration. The trade-offs are significant: each experimental cycle requires multiple LLM calls (generation + critique), so latency and token costs scale linearly with experiment depth; you need a reliable way to execute and measure experiments (simulator, wet lab API, or benchmarked environment); and the knowledge buffer grows unbounded, requiring careful context management for long-running campaigns. For production systems, this works best in R&D phases where discovery speed matters more than real-time response, and where experiments are expensive but infrequent enough that LLM costs are acceptable (e.g., drug screening, materials discovery, hyperparameter optimization).
Limitations and When Not to Use This
The paper's abstract doesn't specify performance metrics, so we can't assess how often EvoMaster's self-critique actually leads to successful hypothesis refinement versus noise—this is a critical gap for evaluating practical utility. The framework assumes access to a reliable experiment execution environment with clear success/failure signals; for domains with noisy, ambiguous, or delayed feedback (clinical trials, long-horizon simulations), the self-critique mechanism may fail to learn correctly. Knowledge accumulation is LLM-dependent: if the LLM cannot reliably parse or generalize from past experiments, the persistent buffer becomes noise rather than signal. The domain-agnostic claim also obscures integration complexity—defining what constitutes an 'experiment' and its observables is non-trivial and domain-specific, potentially negating reusability gains.
Research Context
This work builds on the emerging literature of LLM-based autonomous agents (AutoGPT, ReAct, Toolformer) but specifically targets the scientific discovery use case, which has seen recent success in protein folding (AlphaFold) and molecule generation (GFlowNets). EvoMaster extends prior agent work by making the iterative feedback loop explicit and composable, inspired by human scientific methodology but formalized as an architecture. The research opens directions in meta-learning for agents (can agents learn better strategies for hypothesis generation?) and long-horizon reasoning under uncertainty. It also connects to broader work on chain-of-thought reasoning and self-improvement in LLMs, positioning agentic science as a new benchmark category distinct from QA, code generation, or dialogue.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::