An AI system formulated a hypothesis, designed experiments, ran them, analyzed the results, wrote a scientific manuscript, and submitted it to a peer-reviewed workshop at ICLR — one of the most competitive venues in machine learning. The paper was accepted, scoring above the average human acceptance threshold. This is not a thought experiment. It happened in 2025, and the implications for how scientific discovery is organized, funded, and evaluated are substantial.
The Survey of a New Field
Gridach, Nanavati, Zine El Abidine et al. (2025), in a survey paper accepted at ICLR 2025, provide a comprehensive map of the emerging field of agentic AI for scientific discovery. These systems — capable of reasoning, planning, and autonomous decision-making — are being deployed across chemistry, biology, and materials science to automate literature review, hypothesis generation, experimental design, and data analysis.
The survey categorizes existing systems along several dimensions: the degree of autonomy (from human-directed to fully autonomous), the scientific domain (drug discovery, materials design, synthetic chemistry), and the stage of the research pipeline they address (from idea generation through experimental execution to manuscript preparation). The picture that emerges is not of isolated demonstrations but of a maturing ecosystem: multiple groups worldwide are building increasingly capable research agents, and the infrastructure for evaluating them — benchmarks, metrics, datasets — is developing in parallel.
The critical challenges the survey identifies are not primarily technical but epistemic. How should we evaluate the reliability of AI-generated hypotheses? When an agent proposes an experiment, how do we assess whether the experimental design adequately controls for confounds that a human researcher would recognize? And how do we handle the attribution question — who is the "author" of a discovery made by an AI agent with minimal human guidance?
The AI Scientist v2
Yamada, Lange, Lu et al. (2025) at Sakana AI provide the most dramatic existence proof. The AI Scientist-v2 is an end-to-end agentic system that produced the peer-review-accepted workshop paper described above. The system iteratively formulates scientific hypotheses, designs and executes experiments, analyzes and visualizes data, and autonomously authors manuscripts — including formatting, figure generation, and reference management.
Compared to its predecessor, v2 eliminates reliance on human-authored code templates, generalizes across diverse machine learning domains, and uses a progressive agentic tree-search methodology managed by a dedicated experiment manager agent. The tree search is the key architectural innovation: rather than pursuing a single research direction linearly, the system explores multiple experimental branches simultaneously, evaluating intermediate results to decide which branches to pursue and which to prune.
The system also integrates a Vision-Language Model feedback loop for iterative refinement of figures — the AI reviews its own visualizations and improves them before submission. Three manuscripts were submitted to an ICLR workshop, with one exceeding the acceptance threshold, marking what the authors describe as the first instance of a fully AI-generated paper successfully navigating peer review.
From Digital Labs to Physical Ones
Hartung (2025), publishing in Frontiers in AI, extends the discussion from computational experiments to laboratory automation — the integration of AI agents with robotic systems that can physically execute experiments. This bridging of digital reasoning and physical manipulation is where agentic scientific discovery moves from proof-of-concept to practical impact.
In chemistry and biology, hypotheses often cannot be tested computationally alone. They require synthesizing compounds, measuring properties, observing biological responses. When AI agents are connected to automated laboratory equipment — robotic arms, liquid handlers, spectroscopes, microscopes — the loop from hypothesis to experimental evidence can close without human intervention. The agent proposes a synthesis, the robot executes it, instruments measure the result, and the agent updates its model.
The implications for research productivity are obvious. But Hartung raises deeper questions about the nature of scientific understanding when the discovery process is automated. A human scientist who runs an experiment develops intuitions about the system — tacit knowledge that informs future hypotheses. An AI agent that runs the same experiment extracts data but may not develop analogous understanding. Whether this matters — whether scientific progress requires understanding or merely reliable prediction — is a question that automated discovery forces into the open.
What Changes
The convergence of autonomous paper writing, comprehensive research agent ecosystems, and physical laboratory automation suggests that the question is no longer whether AI can do science but what kind of science it will do. The current generation of systems excels at systematic exploration of well-defined parameter spaces — the kind of research where breadth of search matters more than depth of insight. Whether AI agents can produce the conceptual breakthroughs that reshape fields — the equivalents of natural selection, general relativity, or the double helix — remains an open and fascinating question.