⚛️ AI Scientist: Automating Research

Large-scale AI systems that do research autonomously. We treat invention itself as an engineering problem.

AI Scientist — Automating Research AI Scientist — Automating Research

We develop large-scale AI systems that do research autonomously, towards an AI Scientist. If discovery is a process, it can be studied, systematized, and engineered — as was done with computation (Turing, 1936; von Neumann, 1945) and, more recently, with language (Vaswani et al., 2017; Ouyang et al., 2022). We treat invention itself as an engineering problem.

Engineering needs a decomposition and something to measure. Ours is Karl Popper’s schema of how knowledge grows — problem → tentative theory → error elimination → new problem (Popper, 1972). Each arrow is a component we can build, evaluate, and improve.

Why now

In the past few years, autonomous research has gone from aspiration to peer-reviewed systems, one component at a time:

  • Search with automatic verification. FunSearch (Romera-Paredes et al., 2024) and AlphaEvolve (Novikov et al., 2025) pair a language model that proposes programs with an evaluator that scores them, and iterate at scale. Results include new constructions in combinatorics, a faster algorithm for 4×4 complex matrix multiplication, and scheduling heuristics running in production data centers.
  • Hypothesis generation at scale. Co-Scientist (Gottweis et al., 2026): specialized agents generate, critique, rank (Elo tournaments, simulated scientific debate), and evolve hypotheses, with most compute spent verifying claims against the literature and databases. Tested with researchers at 100+ institutions, it produced experimentally validated leads in areas including drug repurposing, antimicrobial resistance, and liver fibrosis.
  • End-to-end automation. The AI Scientist (Lu et al., 2026) chains ideation, coding, experiments, analysis, manuscript writing, and self-review; a machine-written paper passed first-round review at a workshop of a top machine-learning conference.
  • Prior work as executable agents. Paper2Agent (Miao et al., 2026) turns a paper’s code, data, and workflows into a Model Context Protocol server (Anthropic, 2024) that a coding agent can call — prior work becomes runnable rather than merely readable, and agents built from different papers can be composed on a new question.
  • Self-improving search. Dream-RSI (Zheng et al., 2026) replays a run’s own discovery history as a free simulator to test alternative exploration policies — what to branch on, how much to parallelize, when to stop — and redeploys the best one. The discovery agent stays fixed; the strategy for exploring ideas improves recursively, reaching equal quality with 1.7–2.4× fewer agent calls or generations, and over 50× less budget in one optimization setting.

The ingredients exist, including capable foundation models, agentic scaffolding, tool use, verification, and now meta-level search. What is missing is combining them at the scale, openness, and rigor that real science demands — what has been called the fifth paradigm of discovery (Microsoft Research, 2022), after experiment, theory, simulation, and data.

Engineering the loop

  • Tentative theory → generate. Foundation models make candidate hypotheses, code, and experimental designs cheap. Generation is no longer the main bottleneck; deciding what to generate next is. That decision lives in the search policy and the harness around the model, and both can be optimized in several ways, including by replaying history (Zheng et al., 2026), by letting models optimize their own prompts (Yang et al., 2024; Agrawal et al., 2025), and by end-to-end harness search (Lee et al., 2026) — the harness around a fixed model can move results on the same benchmark by up to 6×.
  • Error elimination → verify. Whatever is verifiable will likely be automated — reinforcement learning from verifiable rewards already drives reasoning ability (DeepSeek-AI, 2025) — so the leverage is in automating verification rather than doing it by hand, including executable tests, reproductions, held-out evaluation, and reward models learned from several kinds of feedback at once (Baur et al., 2026). As in Co-Scientist, most compute should go to checking, not generating. And a verifier must measure what we actually want: solved problems are a proxy for understanding, and optimizing the proxy alone loses the goal (Tao, 2026).
  • New problem → the scientist stays in the loop. A cycle ends with a sharper question, not a finished answer. Domain experts decide which questions matter; the system runs the cycle at a scale and speed no single lab can.

Design choices

  • Large-scale. Research is an iterative, embarrassingly parallel search over hypotheses and experiments, so the system is designed to run and judge thousands of them concurrently — on CSCS’s Alps supercomputer, with the distributed training and inference skills taught in ETH’s Large-Scale AI Engineering course.
  • Open. Scientific claims must be inspectable and reproducible, so the stack is open wherever it can be: open foundation models such as Apertus (Swiss AI Initiative, 2025), open code, and open evaluations, so that results can be checked and built on by others.
  • Grounded. An AI Scientist is only as useful as the problems it is pointed at. We work with domain experts on questions that are well-defined, unsolved, and whose answer changes what people do next, and we measure progress on external, uncontaminated benchmarks rather than our own tasks, including fresh competition mathematics (Balunović et al., 2025) and frontier evaluations such as Humanity’s Last Exam (Phan et al., 2025). Machine-learning research itself is the natural first testbed: a minimal, fully open version of the whole loop already exists — an agent editing a training script against a fixed compute budget and a single validation metric (Karpathy, 2026) — and every step of it is verifiable.

What we look for

Bring us any hard problem, domain expertise, or a check — each feeds the AI Scientist we’re building to solve your biggest problems, and society’s.

  • A hard problem: well-defined, unsolved, with data or a simulator to run against.
  • Domain expertise: you know where your field is bottlenecked and can judge whether an answer is right.
  • A check: compute, funding, or partnership to accelerate the work.

Get in touch.

References

  • Agrawal, L. A., et al. (2025). GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning. arXiv:2507.19457
  • Anthropic (2024). Model Context Protocol. modelcontextprotocol.io
  • Balunović, M., Dekoninck, J., Petrov, I., Jovanović, N., & Vechev, M. (2025). MathArena: Evaluating LLMs on Uncontaminated Math Competitions. arXiv:2505.23281
  • Baur, R., Metz, Y., Gkoulta, M., El-Assady, M., Ramponi, G., & Kleine Buening, T. (2026). MAVRL: Learning Reward Functions from Multiple Feedback Types with Amortized Variational Inference. ICML. arXiv:2602.15206
  • DeepSeek-AI (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948
  • Gottweis, J., Weng, W.-H., Daryin, A., et al. (2026). Accelerating scientific discovery with Co-Scientist. Nature. doi:10.1038/s41586-026-10644-y · blog post
  • Karpathy, A. (2026). autoresearch: AI agents running research on single-GPU nanochat training automatically. GitHub
  • Lee, Y., Nair, R., Zhang, Q., Lee, K., Khattab, O., & Finn, C. (2026). Meta-Harness: End-to-End Optimization of Model Harnesses. arXiv:2603.28052
  • Lu, C., Lu, C., Lange, R. T., Yamada, Y., Hu, S., Foerster, J., et al. (2026). Towards end-to-end automation of AI research. Nature. doi:10.1038/s41586-026-10265-5
  • Miao, J., Davis, J. R., Zhang, Y., Pritchard, J. K., & Zou, J. (2026). Reimagining research papers as interactive and reliable AI agents. Nature. doi:10.1038/s41586-026-11044-y
  • Microsoft Research (2022). AI4Science to empower the fifth paradigm of scientific discovery. blog post
  • von Neumann, J. (1945). First Draft of a Report on the EDVAC. doi:10.1109/85.238389
  • Novikov, A., Vũ, N., Eisenberger, M., et al. (2025). AlphaEvolve: A coding agent for scientific and algorithmic discovery. arXiv:2506.13131
  • Ouyang, L., Wu, J., Jiang, X., et al. (2022). Training language models to follow instructions with human feedback. NeurIPS. arXiv:2203.02155
  • Phan, L., et al. (2025). Humanity's Last Exam. arXiv:2501.14249 · agi.safe.ai
  • Popper, K. R. (1972). Objective Knowledge: An Evolutionary Approach. Oxford: Clarendon Press.
  • Romera-Paredes, B., Barekatain, M., Novikov, A., et al. (2024). Mathematical discoveries from program search with large language models. Nature, 625, 468–475. doi:10.1038/s41586-023-06924-6
  • Swiss AI Initiative (2025). Apertus: Democratizing Open and Compliant LLMs for Global Language Environments. arXiv:2509.14233
  • Tao, T. (2026). A severe misalignment of AI in mathematics. blog post
  • Turing, A. M. (1936). On Computable Numbers, with an Application to the Entscheidungsproblem. Proc. London Math. Soc. doi:10.1112/plms/s2-42.1.230
  • Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention Is All You Need. NeurIPS. arXiv:1706.03762
  • Yang, C., Wang, X., Lu, Y., et al. (2024). Large Language Models as Optimizers. ICLR. arXiv:2309.03409
  • Zheng, T., Wu, X., Zhang, Z., et al. (2026). Dream-RSI: Recursive Self-Improvement through Evolving Worlds. arXiv:2609.14858