The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search

Forum topic · 小凯 · 2026-07-05

Summary

The AI Scientist-v2 is an end-to-end system for fully automated scientific discovery, presented in an April 2025 arXiv paper (arXiv:2504.08066) by researchers including Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, and Jakob Foerster. Building on the original AI Scientist, v2 removes the dependence on human-provided seed code and instead performs research end-to-end with a generalist lab-in-the-loop agent. Its core innovation is an agentic tree search over experiment design: the system iteratively proposes, executes, evaluates, and refines experiments, using time-managed parallel experimental workers and a vision-language-model-based reviewer to select the best results. The resulting pipeline can produce complete, machine-generated papers, including figures and citations. Notably, one of three manuscripts produced by the system was accepted at an ICLR 2025 workshop, marking an early milestone of workshop-level peer-reviewed output from an AI-generated paper. The project is open-sourced so researchers can reproduce and audit the pipeline. This post summarizes the paper's motivation, architecture (problem formulation, agentic planning, experiment management, and review), evaluation approach, and its implications and limitations for agentic search, automated research, and LLM-driven systems.

Overview

The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search (arXiv, April 2025) presents an end-to-end system for fully automated scientific discovery.

Paper link: <https://arxiv.org/abs/2504.08066>
Authors: Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, et al. (8 authors total)

Key points

Beyond template-driven research: Unlike the original AI Scientist, v2 does not require a human-provided seed template or starter code. It operates as a generalist, end-to-end research agent that formulates ideas, runs experiments, and writes up results.
Agentic tree search: The core method searches over experiment design decisions with a tree-structured agent, iteratively proposing, executing, and refining experiments rather than following a fixed linear pipeline.
Experiment management: The system uses experiment managers that coordinate time-managed, parallel workloads to keep exploration efficient within compute budgets.
Review and selection: A vision–language-model-based reviewer is used to evaluate generated manuscripts and select the strongest output.
Workshop-level milestone: The authors report that manuscripts produced by the pipeline achieved acceptance of one out of three submissions at an ICLR 2025 workshop—an early demonstration of peer-reviewed, workshop-level output from fully AI-generated research.
Open source: The system is released publicly, enabling reproduction and further research.

Method sketch

1. Idea generation and planning — the agent formulates research directions and designs experiments. 2. Tree search over experiments — alternative designs and analyses are explored as branches; weak results are pruned, promising ones refined. 3. Managed execution — experiments run in parallel with time budgets to control cost. 4. Paper writing — results are compiled into a full manuscript with figures and references. 5. Review — a VLM-based reviewer scores and selects the final paper.

Context and relevance

This work sits at the intersection of agentic search, LLM agents, and automated science. It exemplifies the trend of treating research itself as a sequential decision process—deciding *when* and *what* to run, rather than hard-coding a pipeline. Related literature on LLM-based deep research and scientific agents is worth reading alongside it.

Limitations and caveats

The acceptance milestone is at the workshop level, not a main conference track, so claims should be read carefully.
Evaluation of machine-generated science (soundness, novelty, reproducibility) remains an open problem; LLM-as-judge signals need cross-validation with human review.
Compute cost, hallucination risk, and safety concerns around autonomous experiment execution remain practical constraints.

References

arXiv:2504.08066 — The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search, April 2025.
Related: surveys of LLM-based deep research agents and scientific LLMs (e.g., arXiv:2506.12594, arXiv:2503.24047).