AI for Science has crossed a quiet watershed. For the last few years the question was: can an AI help a scientist? Literature retrieval, data analysis, code generation, even simulation and experimental analysis — all have been automated one task at a time. The new question is different: can an AI carry an entire research project — from a vague hypothesis to a verified conclusion — instead of completing isolated tasks?
This week, China's CAS-backed lab Zidong Taichu (紫东太初) claims an answer: an upgrade to its ScienceClaw research system called AutoProject, an engine that aims to make "project" the working unit of AI research, not "task." Treat the reported benchmarks with healthy skepticism, but the architecture says something about where the whole AI4S field is heading.
What "project-level" research actually demands
A real research project is a nonlinear exploration loop, not a delivery checklist. It starts from a fuzzy conception, contains interdependent sub-tasks, and constantly reroutes as new papers, experimental results, and anomalous data arrive. Execute every sub-task perfectly and you still haven't done the project — the route itself has to be re-planned mid-flight.
AutoProject's response is a three-layer stack:
- Project2Task — project-level planning. It answers "how should this project be done?": it decomposes a research goal into a task network with dependencies.
- TaskExecutor — long-horizon autonomous execution. It answers "once the direction is set, how do we keep going?": it runs the loop, monitors state, and handles failures without a human babysitting each step.
- EviGraph — evidence-driven verification. It answers "is the final conclusion actually supported?": it builds a traceable graph connecting claims to data, code, and experiments.
The claimed lifecycle: project planning → sub-task decomposition → long-horizon execution → evidence verification → dynamic repair → consolidation of results.
The worked example: a YOLO modeling project
The most concrete illustration is an object-detection modeling task. The system splits it into three modules — data processing, modeling experiments, iterative optimization — and chain-links them: read relevant papers, prepare the dataset, supervise training, handle anomalies, then automatically output metrics, charts, and quantitative analysis.
Where it gets interesting is failure handling. If a task fails or results look anomalous, TaskExecutor retroactively evaluates the step, rebuilds the task network, and reruns experiments with adjusted hypotheses or model parameters — no high-frequency human intervention. Scattered experiment outputs get integrated into a coherent research conclusion rather than a pile of logs.
Underneath sits ScienceClaw's runtime: a multimodal base model (in development since 2021), a hierarchical multi-agent architecture that auto-generates a task graph and dispatches domain, code, search, data-analysis, and simulation agents on demand, plus a cloud-thinking / sandbox-execution / edge-action loop. Coverage is claimed across life sciences (protein evolution being one worked example), materials, chemistry, physics, and astronomy.
The framework: task-level tools vs. system-level science
Strip away the product name and the shift is structural:
- Work unit: from Task (one call, one return) to Project (a closed loop with planning, execution, and verification).
- Capability level: from tool-level (an AI that assists a scientist) to system-level (an AI system that owns the research process and self-corrects it).
- Competition axis: from single-point model or single-agent performance to end-to-end system capability — the ability to keep a long project moving and prove the result at the end.
The analogy that works: a task-level agent is a brilliant intern you hand one assignment at a time; a project-level system is the team lead who turns a vague brief into a plan, re-plans when results contradict assumptions, and comes back with a verified deliverable. Most research agents today are brilliant interns — moving from individual to system intelligence is where agent architecture is heading.
What this means for the AI4S field
- Evaluation is already shifting: yesterday's debate was how to score AI scientists (the evaluation standard is shifting from exams to discovery); tomorrow's is whether a system can drive a multi-week project to a verified conclusion. The reported ARCBenchML composite of 0.865 and above-baseline evidence traceability are vendor numbers — but the metric itself is the real signal.
- Verification becomes a moat: if every lab's model can execute code, the differentiator moves to EviGraph-style evidence graphs — can the system prove, traceably, that its conclusion follows from data it actually handled?
- Commercial vs. open: open research agents (e.g., Polaris-style deployments) are excellent at pipeline automation; project-level autonomy with self-repair loops is where the commercial labs are concentrating — expect this gap to become a talking point.
- Nonlinearity is the bar: any vendor claiming "project-level AI science" should be tested the way AutoProject claims to work: throw it a mid-project contradiction, change the data, and watch whether it re-plans instead of crashing or repeating.
What to do about it
- If you're a researcher: stop benchmarking agents on single-task scores alone. Test the long game: give a system a project, inject conflicts mid-run, and audit whether it re-plans and documents evidence.
- If you're building research agents: steal the three-layer pattern — planning, long-horizon execution with failure loops, and evidence graphs. The self-repair loop is worth more than a slightly better base model.
- If you're buying AI4S tools: ask for the traceability artifact, not the demo. "It finished the project" is a claim; "here is the evidence graph from hypothesis to conclusion" is a deliverable.
FAQ
What is the difference between AutoProject and a typical research agent?
A typical agent completes tasks: one call, one return, and it stops on failure. AutoProject claims to run projects: Project2Task planning → TaskExecutor long-horizon execution → EviGraph verification; on failure it retroactively re-plans the task network and reruns experiments, integrating scattered outputs into a full research conclusion.
Can AI really carry a research project end to end?
Per vendor disclosure: the YOLO example chain-links data processing, modeling experiments, and iterative optimization; ARCBenchML composite is 0.865 with above-baseline evidence traceability. Unverified numbers, but the architecture direction — project-level autonomy plus evidence graphs — is credible.
How do you tell whether a project-level AI science system is real?
Inject a contradictory result mid-project or change the data, and watch whether it re-plans instead of crashing or repeating; then ask for an evidence graph from hypothesis to conclusion, not a demo.
Sources: Zidong Taichu / ScienceClaw AutoProject announcement coverage (QbitAI, August 2026); reported ARCBenchML results are vendor-disclosed and unverified.
