acceptodds
Under review as a conference paper at ICLR 2027

EFiMAS: An Evidence-Guided Multi-Agent System for Task-Conditioned EEG–fNIRS Fusion

Abstract

While EEG and fNIRS provide complementary views of neural processes, the cross-modal relationships of interest vary depending on the task and available measurement information. Existing methods typically rely on predefined computational structures, limiting their ability to adapt fusion design choices to the current task and recorded data. To address this, we investigate EEG–fNIRS fusion as a task-conditioned problem of program construction and revision. We introduce EFiMAS, an evidence-guided multi-agent system. Using dual-index Retrieval-Augmented Generation (RAG), role-specific agents in EFiMAS retrieve neuroscience evidence to propose and independently verify a modeling relationship, while utilizing method evidence to construct and dynamically revise the executable pathway based on observed outputs. Empirically, evaluations on ten public EEG–fNIRS tasks, along with comprehensive component ablations and execution trace analyses, demonstrate that EFiMAS effectively synthesizes task-specific programs and outperforms both dedicated fusion networks and foundation models.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, fNIRS, Multi-agent System, RAG

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_px2XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EFiMAS, a multi-agent system for EEG–fNIRS fusion. Since the relationship between the two modalities is not fully understood and can vary across tasks, especially in terms of temporal and spatial patterns, the authors aim to build task-specific fusion pipelines. The system uses four agents for planning, verification, execution, and review. It also uses RAG to retrieve neuroscience and methodological evidence, which helps the agents propose fusion hypotheses, select appropriate operations, and refine the pipelines based on execution results. Experiments on ten EEG–fNIRS tasks show that EFiMAS achieves competitive performance and outperforms several baselines.

Critical strengths
  1. The paper addresses an important problem, as EEG and fNIRS have different temporal and spatial characteristics and may require different fusion strategies across tasks.

  2. The overall framework is reasonably well designed and clearly presented, with enough implementation details to understand how the system works.

  3. The experiments cover ten EEG–fNIRS tasks, and the results show good performance against several baselines, with additional ablation studies to support the findings.

Critical weaknesses
  1. The novelty is not very clear. Most of the system follows a fairly standard RAG and multi-agent workflow, with limited design specifically for EEG–fNIRS fusion beyond the domain knowledge and tools.

  2. The paper is easy to follow and has a very long appendix with lots of implementation details, but some key questions about the proposed method and its actual contributions are still not well addressed.

Questions for authors
  1. The T/E/H/S relations seem to be one of the main EEG–fNIRS-specific designs, but it is not clear how much they actually affect the generated pipelines. Are they used to guide the search, or mostly just as knowledge for the agents?

  2. There are several ablation studies, but a budget-matched comparison with a single agent seems to be missing. Would a single agent with the same RAG, tools, and budget achieve similar results?

  3. The paper argues that fixed fusion models are not flexible enough, but could task-conditioned Transformers or MoE models also handle different fusion strategies across tasks? There does not seem to be a comparison with these approaches.

  4. The system relies heavily on existing scientific knowledge for hypothesis generation and verification, which seems to favor exploitation over exploration. Could this limit the agents' ability to discover new or unexpected cross-modal relationships?

  5. The Planner proposes hypotheses and falsifiers, but it is not very clear how these are actually tested. Also, could some of the performance gains simply come from having more attempts and execution steps?

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

The authors present an LLM agent pipeline consisting of a planner, verifier, executor, and reviewer. The goal of the pipeline is to assemble an EEG-fNIRS classifier. The pipeline uses separate neuroscience and method RAG while using a library of 227 tools. The output consists of feature extraction, followed by feature fusion and a classifier. For the different agents, they mainly use DeepSeek V4 Pro (but experiment with other models). They provide an extensive evaluation on 10 public datasets and multiple baselines (e.g., fusion networks, EEG foundation models and unimodal networks). The authors claim that EFiMAS reaches the best balanced accuracy (BA) on 7 out of 10 tasks.

Critical strengths
  1. The authors honestly report the negatives in the appendix, such as the loss on ds004514 in section B.1.1 or the up-weighting of corrupted fNIRS in section B.3.4.

  2. The evaluation seems to be done carefully using fixed splits and withholding test labels from all agents. It is further shown that frozen program replay has only a difference of about 0.01 BA.

Critical weaknesses
  1. The improvement of BA over the strongest baseline is a small +0.0256 when compared to other effects. For example, Table 9 shows two "Planner + Executor" runs yielding 0.71 and 0.52 BA on Figshare, and changing the LLM backbone on HYGRIP by 15.8 pp. It is also unclear how statistically significant the BA improvement is, as uncertainty estimates are provided only for EFiMAS (Table 5) and are missing for the baseline result in Table 1. Two out of the seven wins are close to chance (e.g., ds004022 with 0.3 BA, while chance would be 0.25). The baselines also seem to be weak, as it is not clear what kind of hyperparameter search was done and several baselines seem to have collapsed to constant prediction by reporting BA exactly at chance and a macro-F1 value that is that of a constant predictor (e.g., 0.5/0.3333).

  2. Even though several ablations are provided, there is no clear explanation given for why it works, therefore preventing the learning of a transferable lesson. The ablations present only single trajectories, which, given the variance in performance discussed in point one, make differences of 2-10 pp as in Table 2 insignificant. Some ablation results even seem to contradict each other, as removing method evidence hurts in Table 2 but helps in Figure 11. Some ablations are also missing or unclear. While Figure 16 presents the performance increase over the 10 attempts, it remains unclear how much that improves the performance in comparison to other components like RAG, as an ablation separating them is missing. The only no-RAG run I found was in Table 9, but that is without the loop.

Given these points, the main text leaves me with no insight I can reuse or rely on. This is why I recommend rejection. If the improvement can be shown to be statistically significant and similar statistically significant results can be provided on the (missing) ablations, I would raise my score.

Questions for authors
  1. Were the dataset-specific tools (e.g., joint_ds004022_gated_residual_upstream) of the pipeline written by the authors and was this done before or after seeing the results?
  2. What are the uncertainty estimates for Table 2 for multiple independent trajectories? Are the improvements statistically significant?
  3. Why do different independent trajectories differ, i.e., what is the LLM decoding temperature and how else does randomness influence the pipeline?
  4. How much do the 10 validation-guided attempts influence the performance of the pipeline? How much worse or better is a 1-attempt EFiMAS compared to a 10-attempt EFiMAS, and a 10-attempt EFiMAS compared to a 10-attempt EFiMAS without RAG?
Reviewer_HcomOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EFiMAS is an evidence-guided multi-agent system for task-conditioned EEG–fNIRS analysis. It uses separate neuroscience and methodological retrieval indexes, together with planning, verification, execution, and review, to construct task-specific prediction pipelines. The resulting programs combine signal processing, feature extraction, fusion, and classification, and often use relatively simple classifiers. Evaluations on ten tasks report the highest balanced accuracy on seven tasks.

Critical strengths
  1. The framework addresses a meaningful problem: adapting multimodal analysis to differences in tasks, signal representations, and acquisition conditions.

  2. The evaluation covers ten tasks and several recent baselines. The appendix provides substantial additional analyses, including evidence interventions, execution audits, and program replay.

  3. The final task-specific programs are disclosed as executable processing paths, making the outputs more inspectable than agent-generated explanations alone.

Critical weaknesses
  1. The source of the gains remains insufficiently isolated. The tool library already includes specialized feature and fusion operations, while several final programs use conventional classifiers. Comparisons with simpler agents and feature-based search using the same tools, evidence, and budget are needed to establish the contribution of the proposed agent architecture.

  2. Although held-out isolation is explicitly stated, the construction of the research-validation partition and its relationship to the outer evaluation folds are insufficiently specified. Table 2 also uses BA-selected attempts for ablations but a Reviewer-frozen program for Full, making the selection procedures inconsistent.

  3. The component evidence is mixed. Core ablations cover only three tasks, and the role-chain comparison in Table 9 does not consistently favor the full system. Repeated end-to-end program-generation runs are needed to distinguish component effects from search variability.

Questions for authors
  1. How are research-training, validation, and test data separated? Is program selection performed independently within each outer fold, and which partition supplies the BA used to select ablation attempts? Table 2 appears to use different program-selection rules for the ablated and full systems. Please clarify why the same selection rule is not applied to all conditions.

  2. Which task-specific tools and dataset assignments were defined before evaluation, and which operations were newly generated or modified by the agents?

  3. How does EFiMAS compare with a simpler agent or feature-pipeline search given identical tools, evidence, and search budgets? Are baseline tuning budgets comparable?

  4. Can the main component comparisons be repeated across independent program-generation runs using a common selection rule?

  5. The classifiers employed by the agent are rather simple. Why are the more advanced models not selected by the agents? Can they improve the performance?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.