AI for science is moving from "do this task" to "run this whole project"

For most of the last few years, AI for science has meant giving a model a single, well-defined job: predict a protein fold, classify a cell image, run a molecular dynamics simulation. The researcher drove the loop. The AI was a step in it.

A new generation of systems is being designed to take a research goal, phrased in plain language, and run the whole project: decompose it into tasks, schedule the tasks, execute them, gather evidence, check the evidence, and produce a research artifact. The AI stops being a tool. It starts acting more like a junior colleague.

A steel mill wants a vision model that flags surface defects on the production line. In the old model, a researcher would frame the problem, pick an architecture, label data, train, evaluate, and iterate. In the "project-level" model, the user states the goal. The system plans the task graph: split the work horizontally across model variants, then vertically across data and evaluation stages, route dependencies, schedule parallel and serial steps. It then executes the plan, handles exceptions, and emits a metrics report. The human reviews a finished project, not a half-finished pipeline.

That workflow is the headline claim of AutoProject, a research engine released by Zidong Taichu (中科紫东太初), a large-model venture out of the Chinese Academy of Sciences' Institute of Automation. The system sits inside the lab's broader ScienceClaw research agent and is built in three layers: Project2Task for planning, TaskExecutor for long-horizon execution, and EviGraph for evidence-graph verification. The vendor's positioning, carried in 量子位's launch coverage, is that AI4S is graduating from "task agent" to "project agent."

The third layer is EviGraph, an evidence graph: a data structure that links each claim to the experiment that supports it, and checks the chain for consistency. Without that step, the system has no way to tell whether the project has produced a real result or a confident wrong one. On the lab's own benchmark, ARCBenchML, EviGraph posts a composite score of 0.865, against a prior best baseline of 0.596. Its Result Analysis sub-metric moves from 0.442 to 0.794. The Claim Support Rate, meaning what fraction of claims the system can tie to the evidence graph, is reported at 0.38, against a best baseline of 0.27, a roughly 40% relative lift. The Experimental Data Consistency (EDC) score is 0.88.

Those numbers are vendor self-reported. They are not independently audited, and no third-party lab has reproduced them on the same benchmark in the public record. A worked example, a YOLO-based defect detection pipeline that auto-runs data processing, training, exception handling, and emits a metrics and chart bundle, is vendor-curated. A 2025 Xinhua report on the Zidong Taichu 4.0 release and a 2026 China Daily piece on China's 60,000-chip AI4S cluster place the lab in a real national research context, but neither audits the new engine.

A system that owns the whole loop can own the whole mistake. If the planner and the verifier are trained on the same assumptions, they can confirm each other's errors: a confidently wrong intermediate result gets routed into the evidence graph as supporting evidence, the graph checks out internally, and a fabricated finding is delivered to the user as a finished project. Silent bad science, where nothing crashed but the result is wrong, becomes the dominant risk. The audit gap is no longer between the AI and a human reviewer. It is inside the system itself.

The practical test: can a reviewer, or a separate system, walk the evidence graph end-to-end and see exactly which claim was supported by which experiment, without trusting the planner that produced it? If yes, the loop is auditable. If no, the system has automated the production of confident, internally consistent wrong answers. The vendor calls the moment a watershed. Until outside labs run the same benchmark, hold the system to adversarial inputs, and publish the audit, "project-level" is a category the system is reaching for, not one it has crossed.