For centuries, the scientific method has been a fundamentally human endeavour. Researchers conceive hypotheses, design experiments, interpret findings, and communicate results. That paradigm shifted in March 2026 when Nature published research describing The AI Scientist – a system that autonomously navigates the entire research pipeline from ideation to publication. Remarkably, one of its generated manuscripts passed peer review at an ICLR 2025 workshop, marking the first instance of a fully AI-authored paper clearing formal academic scrutiny.
This breakthrough research was co-authored by Vector Faculty Member and Canada CIFAR AI Chair Jeff Clune, alongside researchers from Sakana AI, University of British Columbia, and University of Oxford. The work demonstrates how modern foundation models can be orchestrated into entirely autonomous agentic systems capable of scientific reasoning, experimental execution, and self-evaluation.
TLDR: Uncover groundbreaking AI research in 3 minutes
This concise summary bridges the gap between complex scientific advancements and everyday understanding. Ideal for enthusiasts and non-researchers, start listening now.
System architecture: An end-to-end research pipeline
The AI Scientist operates through four sequential phases: ideation, experimentation, manuscript generation, and automated peer review. Unlike earlier systems that automated narrow tasks – such as discovering chemical structures, proving mathematical theorems, or predicting protein structures – this pipeline handles the complete research workflow.
The researchers evaluated The AI Scientist in two configurations. The initial system, v1, was itself a significant achievement: capable of producing complete research papers by building on human-provided code templates across several machine learning research areas. The template-free v2 represents a further leap from that already capable foundation. Where v1 built on human-provided experiment templates and seed codebases, v2 moves closer to template-free autonomous research: starting from a broad research direction, it can generate and iteratively refine experimental code rather than primarily modifying a fixed human-authored starting template.
Rather than relying on a single general-purpose model, the system orchestrates multiple frontier models across complementary roles, including research ideation, experimental coding, code critique and debugging, vision-language assessment of figures and outputs, and automated peer review.
What the AI scientist actually does
At a high level, The AI Scientist does what a research team does: it reads existing literature, proposes research questions, runs experiments, analyzes results, and writes up findings for publication. The difference is that it does all of this autonomously, without a human doing the work in each step.
Starting from a topic area, the system generates novel research ideas by reviewing existing work and identifying open questions, simultaneously exploring multiple experimental branches. It then writes and executes its own experimental code, analyzes the resulting data, generates figures, and produces a complete manuscript, and even verifies the quality and corrects its own visual outputs using a vision-language model (VLM) feedback loop. Its final output isn’t a rough draft, but a complete, self-assessed scientific paper ready for submission. The pipeline can operate autonomously end to end. Optionally, humans can intervene to select the most promising work at each major stage of the pipeline to speed up efficiency, which the team did when producing the papers eventually submitted for human review.
The peer review test
The most compelling validation came from putting the system to the same test faced by human researchers. With approval from the University of British Columbia’s Institutional Review Board and full cooperation from ICLR 2025 leadership, the researchers submitted three AI-generated papers to the I Can’t Believe It’s Not Better (ICBINB) workshop. Reviewers knew some submissions were AI-generated but not which specific papers, preserving blind review conditions.
One manuscript received an average score of 6.33 – above the workshop’s acceptance threshold. The organizers confirmed it would likely have been accepted had it not been withdrawn per the pre-established protocol. The paper ranked among the top 45 per cent of all submissions that cycle.
The other two papers fell below the acceptance threshold. The AI Scientist cannot yet consistently produce workshop-quality work and none of the submissions approached the bar for main conference acceptance, but the system produced a paper that passed peer review at a top-tier machine learning conference workshop – and that is a meaningful line crossed.
A scaling law for scientific quality
Understanding the broader significance of these results required a way to evaluate AI-generated papers at scale, which is costly for human reviewers to do. The researchers developed an Automated Reviewer that ensembles five independent LLM judgements into a meta-review, then uses a final model to synthesize a consensus, mirroring the area chair role at major conferences. Validated against real-world ICLR submission data, the Automated Reviewer predicted acceptance decisions with 69 per cent balanced accuracy – similar to the 66 per cent balanced agreement rate observed between human expert reviewers in a comparable study, giving the research team a credible, scalable proxy for human expert judgment.
When researchers used this tool to compare paper quality across different foundation models ordered by release date (a good proxy for the capability of the model, since later models are more capable), they found that newer, more capable models produced much better papers (the trend is clear and significant). Similarly, more compute allocated to the system’s experimental search also produced better papers.
Together, these findings suggest a scaling law for scientific quality: as the underlying models improve – and they are improving rapidly – so does the science they produce. Relatedly, we can produce higher quality by allocating more compute to The AI Scientist. The AI Scientist as it exists today is thus a floor, not a ceiling. This trajectory points beyond incremental gains and toward a fundamental shift in how research itself is conducted.
Towards a second scientific revolution
The implications of this work extend well beyond machine learning. Science has always been constrained by human bandwidth: the number of hypotheses a researcher can test, the experiments a lab can run, the papers a community can produce and review in a year. Those constraints have shaped the pace of discovery for generations.
A system like The AI Scientist begins to challenge those constraints directly. If AI can autonomously run experiments, analyze results, and produce peer-reviewable papers, the pace of scientific discovery need not be tied to the number of trained researchers available to do the work. Run in parallel across many instances, such a system could explore research directions simultaneously that would take human communities decades to cover.
The work invites a comparison to the first scientific revolution – the shift in the 16th century from natural philosophy to empirical science. That transformation was not just about new discoveries, but a new method for making them. The AI Scientist suggests we may be entering another such shift: one where the method is augmented by autonomous systems capable of running the scientific process at scale.
This does not make human researchers redundant. It just redefines their role. The researcher’s focus shifts from executing experimental pipelines to directing them – asking the right questions, identifying surprising results, and exercising the judgment to know what matters. Researchers who can effectively direct AI systems toward meaningful scientific questions may be able to pursue programs of much greater scope than those available to any individual today.
The near-term implications are most concrete in fields where experiments run computationally, such as machine learning. But other fields that harness computer models to accelerate discoveries and/or have increasingly more capable automated laboratories where robots can conduct experiments will also be transformed, such as drug discovery, synthetic biology, chemistry, materials science, and climate modelling. For fields requiring more difficult physical experimentation, the path is longer.
Responsible development
The current system has real limitations. Papers it produces are not uniformly good: common failures include underdeveloped ideas, incorrect implementations, experimental errors, and hallucinated citations. No AI-generated paper has yet cleared the bar for main conference (or top-tier journal) acceptance, and The AI Scientist currently only operates within computational domains.
The broader risks are significant and the researchers are direct about them. A system capable of producing plausible-sounding papers at scale could overwhelm peer-review processes already under strain or add noise to the scientific literature that proves difficult to filter. The research team obtained explicit IRB approval and withdrew all submissions post-review to avoid setting norms before the community is ready to establish them. There are also other risks, such as systems that conduct dangerous research or create powerful, yet unaligned AI.
These challenges are not arguments against the research, but for developing it carefully – building the disclosure standards, evaluation frameworks, alignment and oversight techniques, and community norms now, while there is still space to do so thoughtfully.
Read the full paper
This post covers the key ideas, but the full Nature paper contains extensive implementation details, ablation studies, supplementary analysis, and more detailed discussions of limitations and safety issues. The AI Scientist-v2 technical paper provides additional detail on the template-free architecture. Open-source implementations of both systems are available on GitHub.
While an early system, The AI Scientist demonstrates that the complete research pipeline is now within reach of autonomous AI. Where that leads is one of the more consequential open questions in science right now.
Created by AI, edited by humans, about AI
This blog post is part of our ‘ANDERS – AI Noteworthy Developments Explained; Research Simplified’ series. Here we utilize AI Agents to create initial drafts from research papers, which are then carefully edited and refined by our humans. The goal is to bring you clear, concise explanations of cutting-edge research conducted by Vector researchers. Through ANDERS, we strive to bridge the gap between complex scientific advancements and everyday understanding, highlighting why these developments are important and how they impact our world.
