AGTBRIEF
UNC's CyberNeuro preprint — a four agent, locally run AI workbench for brain scans — lifts test set accuracy from 40% to 69% on NeuroBench, a public brain imaging benchmark, by running on a small local model called WandaMind, while a cloud LLM
A junior researcher at a small brain-imaging lab might spend a week hand-curating scan metadata for a single study, babysitting a pipeline that drops steps at random, and checking every output by eye for a misaligned slice. A new preprint from UNC Chapel Hill's ACMLab argues that an on-site AI workbench can take that work over without sending patient scans to a cloud model.
The workbench, CyberNeuro, splits the job across four cooperating LLM agents: a planner, validator, dispatcher, and reporter that talk through a secure local bridge. The default brain behind them is WandaMind, a tailored local model. Sensitive scan data stays on the lab's machine. A cloud LLM stays available only as a comparison lane.
On the public NeuroBench suite, the authors report held-out domain accuracy rising from 40% to 69% over their baseline. Across a 10-batch end-to-end workflow, the local WandaMind setup used about 10.6% of the tokens, and a cloud-provider config about 61.7%, both measured against the same paper-defined baseline.
Both numbers are author-reported. The token share is relative to a baseline the paper calls "Neuroclaw," not a generic cloud LLM. No third-party NeuroBench placement or clinical-site deployment is on the public record yet. The team has published the platform and a demo video, and the preprint defines the four agents in full.