acceptodds
Withdrawn from ICLR 2027

Towards a Virtual Neuroscientist: Autonomous Neuroimaging Analysis via Multi-Agent Collaboration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Transforming neuroimaging data into clinically actionable biomarkers is a knowledge-intensive and labor-intensive process. Standardized workflows such as fMRIPrep have improved robustness and efficiency, but they are statically configured and cannot reason about downstream objectives, deliberate over alternative strategies, or close the loop between intermediate evidence and subsequent decisions in the way a human researcher would. This lack of closed-loop adaptation often leaves domain experts trapped in a cycle of manual trial-and-error to tune parameters and remediate pipeline failures, severely constraining the scalability of clinical biomarker development. To bridge this gap, we introduce NEXUS, an autonomous multi-agent framework that integrates neuroimaging workflow execution with scientific-objective understanding. Unlike conventional flat tool-calling agents, NEXUS adopts a code-centric execution paradigm where specialist agents collaboratively synthesize and optimize executable programs over composable domain-specific primitives. This design enables robust, long-horizon workflow construction that adapts dynamically to runtime observations. Furthermore, we propose a hierarchical verification framework for autonomous quality control, integrating cohort-level metric screening with agentic visual inspection to drive evidence-grounded workflow remediation. Experiments on ADHD-200 and ADNI demonstrate that NEXUS outperforms automated workflow-based baselines in predictive performance while exhibiting sophisticated agentic behaviors, including strategy exploration and adaptive refinement.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neuroimaging Analysis, Multi-Agent, Scientific Workflow

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_e47pOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary
  • The paper presents NEXUS: an agentic framework for specific types of MRI and fMRI analysis.
  • NEXUS uses pre-defined primitives that wrap preprocessing and analysis tools, and assigns different parts of the task to specialized agent roles.
  • One of these roles does quality control using a VLM, prompted with written criteria for how images should look and a few examples of good and bad cases.
  • Evaluation covers two datasets, with 5 runs each.
  • Baselines are standard tool pipelines, with cross-validation used for hyperparameter tuning and pipeline/model selection.
  • Ablations show each component contributes.
  • The QC agent’s agreement with human raters approaches the agreement between the raters themselves.
Critical strengths
  • Improves upon software used by neuroscientists for image analysis by helping with parameter setting
  • Evaluates how well a pre-trained VLM with prompts and zooming tools can help an agentic framework for automatic pipeline construction
Critical weaknesses
  • The paper does not compare NEXUS to any other agent-based system. This is its most critical weakness.
    • Missing comparisons include both the neuroimaging agents discussed in Related Work and general-purpose coding agents such as Claude Code or Codex, which would be easy to run as baselines.
    • The authors acknowledge the closest works: “During the writing of this work, we noticed several recent works, including NeuroClaw (Wang et al., 2026), NEURA (Xie et al., 2026), Agentic LLMs (Erdur et al., 2026), NeuroAgent (Zhong et al., 2026), NeuroPilot (Chen et al., 2026), and BrainPilot (Li et al., 2026), which share a similar spirit with us in exploring agentic AI for neuroimaging research.” They distinguish NEXUS as follows: “They primarily focus on task completion through tool orchestration, reproducible execution, and traceable workflows, typically instantiating predefined or skill-encapsulated pipelines. In contrast, NEXUS targets fully autonomous, end-to-end neuroimaging analysis, emphasizing adaptive workflow construction, dynamic refinement, objective-aware optimization, and closed-loop agentic QC.” These descriptions sound quite similar to me. The paper should include a quantitative comparison with these systems, or explain why such a comparison is not possible.
    • A comparison with general-purpose agents is especially important. BrainPilot, which the authors cite, compares its custom multi-agent system against Claude Code and Codex. It reports that BrainPilot “attains performance comparable to state-of-the-art agent framework with less costs.” Simply giving the task to Claude Code or Codex is therefore a straightforward baseline, and it is missing here.
  • Some neuroimaging tools discussed in Related Work (e.g., FreeSurfer, BrainSuite, Nipype) are not used in either the baselines or the primitives.
  • To me, the most interesting part of the paper was the VLM-based QC agent. It is not compared to prior automated QC methods, either VLM-based (e.g., MMRQA) or established non-VLM tools (e.g., MRIQC’s classifier), and Related Work does not discuss them.
Questions for authors

Can you provide comparisons to other agentic frameworks?

Reviewer_2M5aOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents NEXUS, a multi-agent framework for automating end-to-end neuroimaging analysis. The system uses a supervisor agent to coordinate specialized agents responsible for data understanding, preprocessing, quality control, and downstream analysis. Its main contributions include a code-centric execution framework that allows agents to construct workflows using domain-specific tools, just-in-time context injection to provide task-relevant tool information, and a hierarchical quality control mechanism that combines quantitative metrics with agent-based visual inspection. The authors evaluate NEXUS on two neuroimaging datasets, ADHD-200 and ADNI, and show that it achieves better downstream prediction performance than conventional automated workflows. They also conduct ablation studies to examine the impact of different system components on execution reliability and evaluate the quality control module through its impact on prediction performance and agreement with human assessments.

Critical strengths
  1. Practical and meaningful problem: The paper addresses an important challenge in neuroimaging research, where constructing and maintaining analysis pipelines requires substantial domain expertise and manual effort. Automating this process could make neuroimaging analysis more accessible and efficient.

  2. Comprehensive end-to-end framework: NEXUS covers the entire neuroimaging analysis process, from raw data understanding and preprocessing to quality control and downstream predictive modeling. This makes the framework more practically useful than systems that automate only individual stages.

  3. ntegration of quality control into workflow decisions: Instead of treating quality control as a separate reporting step, the framework incorporates QC feedback into subsequent processing decisions. This is a meaningful design choice because poor-quality neuroimaging data can directly affect downstream analysis.

Critical weaknesses
  1. Limited comparison with existing neuroimaging agents: Although the paper discusses several recent neuroimaging agent systems, such as NeuroClaw, NeuroPilot, and BrainPilot, the experimental comparisons focus primarily on conventional preprocessing pipelines and exhaustive workflow search. Without direct comparisons to existing agent-based approaches, it is difficult to assess how much NEXUS advances the current state of the art.

  2. Lack of statistical significance analysis: Some improvements over the strongest baselines are relatively small, particularly on ADNI, where NEXUS achieves an AUC of 93.00 compared to 92.55 for exhaustive search. Without statistical significance tests or confidence intervals for these differences, it is difficult to determine whether the improvements are reliable.

Questions for authors

Could the authors provide more evaluation of the intermediate stages of the generated pipelines? It is unclear how well these stages perform and which contribute most to the final improvements.

Reviewer_bxj1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes NEXUS, an agentic framework for human functional neuroimaging workflows. It adopts a code-execution paradigm where multiple specialist agents can collaborate to improve programs. A supervisor agent takes a user-provided query and decomposes it for different agents. The authors conduct experiments on two datasets, ADHD-200 and ADNI.

Critical strengths

NEXUS addresses the important challenge of tackling neuroimaging workflows, which are heterogenous, computationally expensive, and require human domain expert intervention.

Critical weaknesses

The paper makes strong claims about NEXUS being autonomous. The supervisor does perform task decomposition, but the user must supply the actual scientific objective and specifies the target, representation, desired output, etc. So NEXUS demonstrates a degree of workflow following, but misses the critical, scientifically relevant task of autonomously generating scientific questions, hypotheses, or goal formation.

The novelty of JIT context injection seems unclear. This appears to be very closely related to standard tool retrieval and skill routing. There is no comparison with standard approaches (BM25, etc), and only uses the weak comparison of entire primitive library.

The paper states proposing hierarchical agent-as-judge framework for closed-loop autonomous QC as a methodological contribution, but does not distinguish the novelty of its proposed architecture from existing agentic systems that contain critic-revision loops and active visual search methods.

The paper lacks the mandatory AI-use statement for a submission. There is no mention of code availability. Furthermore, for a paper with a focus on human brain datasets, with a downstream goal being the analysis of the data therein, a complete lack of ethics discussion in the main paper is concerning, especially where the paper mentions the pursuit of "clinically actionable biomarkers".

Questions for authors

Clearly justify autonomy of the system.

How is JIT context injection technically different from widespread existing approaches and does it improve on real baselines.

What do the +/- values in Table 1 mean?

Figure 2. For every architectural ablation report downstream performance and quality, not just errors.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.