aclanthology.orgSource publication:
Loading content…
Pretrained neural-recording models differ in architecture and input format, so labs configure each one by hand before they can compare them. NS-Copilot puts nine EEG and spike models behind one natural-language interface, with planner, coder, controller and interpreter agents that select models, generate code and close the loop. Over 8 trials each on Alzheimer's, Parkinson's and working-memory benchmarks, its primary metrics were 57.16% macro F1, 71.32% balanced accuracy and 74.43%, above the listed baselines.
Model selection and invocation for neural data analysis are handed to a set of role-specialized AI agents, with the user supplying only raw recordings and a task description. EEG pretrained models each demand their own sampling rate, channel layout and tensor shape, so cross-model comparison required manual per-model configuration, and general AutoML search spaces do not include these domain models. The system ran 8 independent end-to-end trials on each of three public datasets, Alzheimer's EEG, Parkinson's EEG and working-memory spikes, scoring 57.16% macro F1, 71.32% balanced accuracy and 74.43% on the primary metric, above AutoML-Agent, LAMBDA, AutoGen, SciAgents, Claude-Sonnet-4.5 and Codex baselines.
The nine pretrained models are abstracted as stateless forward-inference endpoints that only emit feature vectors, while downstream classification logic is written at runtime. A planner agent reads modality, channel count, sampling rate and class distribution, sets the primary metric itself and builds a prioritized model queue that stops early once a performance threshold is passed. If the queue is exhausted below threshold, a controller diagnoses execution metrics and routes between targeted code patching and full replanning, with a default of two retry rounds; on Alzheimer's, two forced retries moved the mean by −0.11 points, while returning the best of three rounds was +3.80 points over the initial round, improving 25 of 48 runs and worsening none.
The gain comes mainly from the pretrained models rather than from multi-agent orchestration. The same architecture without pretrained models scored 49.80 and 64.52 on Alzheimer's and Parkinson's in its initial round, below single-step Codex at 51.01 and 65.52. In a held-out evaluation with queue search, early stopping and the controller switched off and the model fixed in advance, REVE beat handcrafted features by 13.34 points on Alzheimer's, POYO-1 by 6.82 on working memory, and BrainOmni by only 0.68 on Parkinson's.
The demonstrated scope is EEG and extracellular spike analysis on three public datasets: three-class Alzheimer's classification, binary Parkinson's classification and left-versus-right lick decoding in working memory, one dataset per scenario. It suits researchers who want to compare several neural pretrained models without writing pipeline code. The authors plan more datasets within each scenario and extension to fMRI, MEG and additional analysis tools.
The authors note that each scenario is represented by a single dataset and that closed-loop evaluation raises execution time and compute, with per-run cost roughly 1.5 to 2 times the single-step Codex baseline. The uniform 300-second script budget made BrainOmni, the model with the largest embedding dimension, time out repeatedly on Alzheimer's and enter the mean as zeros; re-run with a more generous allocation it reached 62.32% macro F1. BrainOmni's checkpoint had used a 65-subject two-class subset of that dataset as its own downstream benchmark, while these experiments freeze the backbone and use all 88 subjects with three classes.