acceptodds
Withdrawn from ICLR 2027

BrainState: Graph-based State-Aware Generalization for Non-invasive Brain-to-Text

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

EEG-based brain-computer interfaces (BCIs) offer a promising pathway for neural language decoding and human-machine interaction. Yet, their generalization is limited by heterogeneous distribution shifts across recordings. Existing transfer learning methods often seek task-relevant representations that are less sensitive to physiological and recording differences. Nevertheless, state-dependent variation can persist in non-stationary EEG and affect how information is transferred across subjects. To address this, we propose BrainState, a human state aware state-space-based hierarchical graph transfer learning framework for EEG-to-language. BrainState first combines spatiotemporal convolutions with Mamba blocks to capture local EEG patterns and temporal dependencies, mapping each trial to task-relevant content and recording-state representations. It then organizes training examples into a hierarchical graph memory according to class, subject, and session, enabling condition-specific modeling and state-aware retrieval. Finally, an attention-weighted memory readout aggregates relevant prototypes, while compatibility-based retrieval and gated fusion limit contributions from poorly matched recording conditions. BrainState thereby provides a way to incorporate recording state into cross-subject EEG speech decoding. Comprehensive experiments demonstrate the SOTA performance of the framework against strong baselines under cross-subject settings. Our code is available online at https://anonymous.4open.science/r/BrainState.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain-Computer Interface (BCI), EEG-to-text decoding, transfer learning, graph neural nwtworks

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_GiAYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes BrainState, a cross-subject EEG decoding framework that integrates the convolutional mamba framework with a memory module. The training trials are organized into a hierarchical memory graph. A state-conditioned retrieval mechanism retrieves the evidence from poorly matched recording then adds them to the classifier through gated fusion.

Critical strengths
  1. The evaluation uses strict leave-one-subject-out splits with per-subject results.
  2. This paper explores an important problem in EEG decoding.
Critical weaknesses
  1. This paper does not compare with the latest brain foundation models, e.g., CBraMod[ICLR 2025], CSBrain [Neurips 2026], CodeBrain [ICLR 2026].
  2. The novelty of this paper is incremental by integrating a memory retrieval mechanism with CNN+Mamba. The motivation of this paper is also not clearly introduced.
  3. The paper claims to resolve eeg-to-language/brain-to-text decoding task. However, the authors only experiment on one dataset named TESSCCo, which involves five TV-control commands. So the task is actually a 5-class task. How does this method perform on more commonly-used text decoding datasets, such as ZuCo [1]?
  4. The performance is limited. The 5-class accuracy is 29.9%, only about 10 percentage points above chance (20%).
  5. The ablation study is conducted on only three of the 21 subjects. This makes the ablation results less reliable and insufficient to demonstrate the effectiveness of state-conditioned retrieval.
  6. The paper lacks qualitative analysis of the retrieval mechanism, for example which training sessions are retrieved for a held-out trial and whether they are actually close in recording state.
  7. The writing needs substantial improvement, and the paper is difficult to follow.

[1] ZuCo, a simultaneous EEG and eye-tracking resource for natural sentence reading, Scientific Data.

Questions for authors

Please see weaknesses.

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

This paper proposes BrainState, which separates EEG representations into content and recording-state components and performs state-conditioned retrieval over a hierarchical prototype memory. It reports improved cross-subject performance on five-class silent-speech command classification using the native Spanish subset of TESSCCo.

Critical strengths
  1. The main idea is intuitive: instead of removing recording-state differences, the paper uses the state to decide which training examples should be retrieved. The EEG-to-language problem is also well aligned with ICLR topics.
  2. The presentation is clear and self-contained. Besides the main evaluation, the manuscript provides a broad set of analyses, including subject-shift, class-separation, and ablation studies.
Critical weaknesses
  1. The methodological novelty is insufficiently established against closely related prototype methods. For example, MAT has separated domain and class features and used domain features to select a relevant prototype space at inference. BrainState 's advantages over MAT are neither directly evaluated nor sufficiently distinguished. Earlier EEG work by Musellim et al. (2022) also combines subject-style modeling with semantic prototype learning.

  2. The experiments do not sufficiently isolate the benefit of state-conditioned retrieval. The core ablations cover only three of the 21 held-out subjects, and the progression to the full model changes both retrieval and training objectives. Although Appendix H includes auxiliary-loss weight sweeps, we still need a comparison that keeps the same encoder, auxiliary losses, memory, and fusion while removing or shuffling the state-matching signal. The additional benefit of the hierarchical organization over a simpler prototype memory also remains unclear.

  3. The performance gain over StateMamba is relatively modest.

Questions for authors
  1. How does BrainState differ from MAT’s domain-conditioned prototype inference, and what improvements remain when both methods use the same EEG backbone and evaluation protocol? Could the characterization of MAT as content-only be revised?

  2. Please provide additional experimental results to verify Weakness 2.

  3. Can the authors provide evidence on another speech EEG dataset and compare against fine-tuned foundation-model baselines to establish the scope and robustness of the improvements?

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

This paper proposes BrainState, a state-aware transfer learning framework for cross-subject EEG command classification. A convolutional–Mamba encoder produces content and recording-state representations. The state representation is trained to reconstruct signal statistics, while additional objectives encourage separation between content and state. The model retrieves evidence from a hierarchical class–subject–session memory using content similarity penalized by state distance, and uses the retrieved evidence to refine its predictions.

The main evaluation uses the Spanish covert-speech subset of TESSCCo under subject-disjoint leave-one-subject-out evaluation. BrainState achieves 29.94% accuracy and 28.21% macro-F1, outperforming the evaluated baselines in mean performance (Table 1, p. 6). Ablations provide initial support for the benefits of learned state representations and the complete framework (Table 3, p. 8).

Critical strengths
  1. A well-motivated approach to recording heterogeneity. The analyses in Figure 4 (p. 7) and Figure 8 (p. 17) demonstrate substantial subject-related structure in the recordings. Using a signal-derived recording-state representation to guide the selection of training evidence is a reasonable response to this problem.

  2. Encouraging results under cross-subject evaluation. The LOSO protocol separates training, validation, and test subjects (p. 5). Table 1 (p. 6) reports improvements over the evaluated baselines, while the fixed-state and learned-state variants in Table 3 (p. 8) provide initial support for learning a state representation rather than relying only on fixed statistical descriptors.

Critical weaknesses
  1. Limited evidence for generalization and limitations of baseline adaptation. The main evaluation is restricted to 1,792 Spanish covert-speech trials from a single dataset and a five-class command classification task (Appendix B.1, p. 15). This limits the evidence that the improvements extend beyond the characteristics of this corpus.

Moreover, Appendix J (p. 30) states that LaBraM and EEGMamba use frozen pretrained encoders, whereas BrainState is trained end to end. The reported comparison therefore does not establish an advantage over suitably fine-tuned foundation models.

  1. Insufficient validation of retrieval calibration and memory structure. Equation (3) (p. 4) combines temperature-scaled cosine similarity with a dimension-normalized squared state distance. However, λ = 1.0 and τ = 0.2 are fixed without a reported sensitivity analysis of these retrieval parameters (Table 9, p. 26). Figure 13 (p. 27) varies auxiliary loss weights rather than these parameters. Although the projection heads include LayerNorm (Table 8, p. 25), the effective balance and variability of the two score components remain unclear.

The ablations also cover only three subjects (Section 4.4, p. 8) and do not separately isolate the benefits of the hierarchy and residual representation in Equation (2).

Questions for authors
  1. What motivated restricting the main evaluation to the Spanish covert-speech subset? How were the frozen-encoder adaptation protocols selected, and were fine-tuning alternatives considered? Any available results on another relevant dataset or a suitably fine-tuned baseline would help assess the scope of the reported improvements.

  2. How were \lambda and \tau selected, and how sensitive are performance and retrieved-node rankings to them? Please clarify the empirical scales and variability of the two score components. Quantifying how often the state term changes the selected nodes, and whether those changes improve predictions, would help establish its practical contribution.

  3. What evidence isolates the benefits of the hierarchical organization and residual representation beyond adding a memory-based prediction pathway? A controlled comparison with flat memory and direct condition-prototype retrieval would help clarify their necessity.

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

Related papers

Discussion (0)

Sign in to comment.