Automated AI Researcher

An automated AI researcher connects models, tools and execution environments across parts of a research process. Literature review, hypothesis generation, experimental coding, execution, analysis and writing are distinct capabilities. AI Scientist-v2 ↗ is an example of computational research automation using agentic tree search.^[1]

AI for Science covers AI applications in science broadly; automated research concerns which research stages a system performs. Using a property predictor does not establish automated topic selection or validation of a discovery.

Sources were checked on 20 October 2026. The 25 March Nature article ↗ describes AI Scientist research.^[2] Publication of a system-description paper and validation of each system-generated manuscript are different evidence.

The 13 October Agentic Scientist preprint ↗ connects task drafting, ideation, execution, evidence generation and writing.^[3] Its project report ↗ describes human direction and verification of implementations, results and citations in the ALIFE case.^[4] This does not establish independent discovery without people across all scientific fields.

ORBiS Tree proposes starting with a question, permitted data, compute budget and comparator, then specifying what observation would count against the hypothesis. Execution produces code and data that actually test it; analysis determines the supported scope; writing derives claims from recorded evidence.

Trace a conclusion backward through figures, run records, implementation and the question. Failed or invalid experiments must not silently become evidence for later stages. A long report without this chain is inadequate. Agentic Coding supports implementation; Data Provenance describes data and transformations.

Suppose twelve ideas yield four implementations, three satisfy predefined methodological conditions, and two retain their result in an independent rerun. The retained-result rate is 2/12≈16.7% over all ideas or 2/3≈66.7% over valid experiments. Both calculations are correct but answer different questions. Showing only four implementations hides the other eight candidates' costs and rejection reasons.

These are teaching figures, not system performance. Define what was independent about the rerun, including code, data and environment. Re-execution on the same data differs from replication in a new population. Neither rate alone measures scientific importance.

Check whether a score-improving change preserves the proposed mechanism. Match baseline data and budgets, and use ablations that remove relevant components. Separate candidate-selection data from final evaluation and inspect whether candidates can modify the evaluator. These are ORBiS Tree's evaluation proposals.

Benchmark Data Contamination addresses selection/evaluation overlap; Hallucination (AI) addresses unsupported generated claims. A manuscript score, external acceptance and correct empirical findings are distinct validation stages.

The Agentic Scientist project's limitations distinguish AI reviewer scores from independent evidence of scientific correctness, and the externally reviewed Error Diffusion paper from internal evaluation manuscripts.^[4] Transfer from computational machine learning to instrument-based, physical or human-subject research needs separate evidence.

Record responsibility for question selection, inappropriate comparisons, attribution and final claims. A finished manuscript does not establish a scientific finding. Read methods, failures and evaluation scope together when judging autonomy.

Read AI for Science for applications, Agentic Coding for code validation and AI Agent Oversight for execution scope. Improving a research output differs from improving the research system itself; the latter is addressed by Recursive Self-Improvement. Model Context Protocol connects data and tools without replacing scientific validation.