Writing study: MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding
← NeurIPS 2026 oral writing corpus
Zexuan Chen, Sichao Liu, Runhao Lu, Huichao Qi, Alexandra Woolgar, Xi V Wang, Lihui Wang · Official NeurIPS 2026 oral record
Figure, table and section ledger
Editorial contribution type: Architecture & representation (primary).
Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.
- Canonical PDF
- 20 physical pages
- Before references / appendix
- ≈ 10.09 page-equivalents
- Main figures / tables
- 5 / 6
- Attached appendix figures / tables
- 5 / 9
- Attached appendix span
- ≈ 5.91 page-equivalents
Source-bound correction: Source-checked 2026-10-02: retain actual Appendices A-D beginning on page 15. Remove two spurious standalone Appendix tokens on pages 19 and 20; rendered pages show only continuation figures and captions there, not section headings.
Section-by-section lengths and source pages
The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.
| Top-level heading | Starts on PDF page | Estimated span | Share of entire PDF |
|---|---|---|---|
| 1 INTRODUCTION | 1 | ≈ 1.48 pp. | 7.4% |
| 2 RELATED WORKS | 3 | ≈ 0.59 pp. | 2.95% |
| 3 METHODS | 3 | ≈ 2.51 pp. | 12.55% |
| 4 EXPERIMENTS AND RESULTS | 6 | ≈ 3.54 pp. | 17.7% |
| 5 Conclusions, Limitations, and Future Work | 9 | ≈ 0.38 pp. | 1.9% |
| Acknowledgments | 10 | ≈ 0.11 pp. | 0.55% |
| Reproducibility Statement | 10 | ≈ 0.75 pp. | 3.75% |
| References | 11 | ≈ 4 pp. | 20% |
| A Dataset | 15 | ≈ 0.34 pp. | 1.7% |
| B Preprocessing and Implementation | 15 | ≈ 0.66 pp. | 3.3% |
| C Additional Experimental Results | 16 | ≈ 2 pp. | 10% |
| D Supplementary Visualizations | 18 | ≈ 2.91 pp. | 14.55% |
All detected figure and table locations
Main text
Argument and organization
- Problem and contribution
The introduction identifies low signal-to-noise ratio and subject variation as obstacles to direct EEG–image alignment, then motivates masked pretraining and tri-modal alignment with generated descriptions. It defines the downstream claim as retrieval among unseen visual concepts, not unconstrained reconstruction of arbitrary thoughts.
- Argument sequence
The method separates the EEG encoder, masked reconstruction pretraining and contrastive alignment; Algorithm 1 connects these stages. Experiments distinguish in-subject and cross-subject protocols before probing pretraining, encoder components and alignment weights, with per-subject results moved to the appendix.
- Evidence and verification
The result prose reports the cross-subject setting separately and retains a non-significant comparison rather than labeling every result significant. The pretraining ablation notes that absolute differences fall within the reported standard deviations and selects a small decoder partly for efficiency; component ablations do not by themselves establish strong causal neuroscientific interpretations.
Problem, claim and contribution
The introduction identifies low signal-to-noise ratio and subject variation as obstacles to direct EEG–image alignment, then motivates masked pretraining and tri-modal alignment with generated descriptions. It defines the downstream claim as retrieval among unseen visual concepts, not unconstrained reconstruction of arbitrary thoughts.
Introduction pp. 1–3; framework §3, pp. 4–5; data protocols Appendix A, p. 15 · Main paper
Argument and section order
The method separates the EEG encoder, masked reconstruction pretraining and contrastive alignment; Algorithm 1 connects these stages. Experiments distinguish in-subject and cross-subject protocols before probing pretraining, encoder components and alignment weights, with per-subject results moved to the appendix.
§§3.1–3.3, pp. 4–5; Algorithm 1 and setup p. 6; results/ablations pp. 7–9; Appendix C, pp. 16–17 · Main paper
Figures, captions and visual explanation
Figure 1 uses modality-specific shapes to show EEG, image and text alignment, then places top-k retrieval examples beside separate in-subject/cross-subject bar charts. Table 1 on p. 7 keeps per-subject MEG columns and protocol blocks; Tables 2–3 use compact ablation layouts with means and standard deviations instead of only a best configuration.
Figure 1, p. 2; Tables 1–3, p. 7; rendered pages 2 and 7 · Main paper
Tables, comparisons and evidence
The result prose reports the cross-subject setting separately and retains a non-significant comparison rather than labeling every result significant. The pretraining ablation notes that absolute differences fall within the reported standard deviations and selects a small decoder partly for efficiency; component ablations do not by themselves establish strong causal neuroscientific interpretations.
Cross-subject result discussion and Tables 1–3, p. 7; qualitative/neural analyses pp. 8–9 · Main paper
Appendix and reproducibility
The appendix defines the unseen-concept and subject splits, gives preprocessing and visual-encoder conditions, and exposes subject-level variability. Additional encoder comparisons and prompt details help distinguish the contribution of the EEG design from different pretrained visual features or text supervision.
Appendices A–B, p. 15; Appendix C and Tables 9–15, pp. 16–17; Appendix D, pp. 18–20 · Main paper
Clarity and writing conventions
The most transferable writing practice is maintaining separate names for in-subject, cross-subject and unseen-concept generalization. The main text's explicit exception to a significance claim and its acknowledgement of overlapping variability make the conclusion more precise than unrestricted state-of-the-art language.
Protocol definitions pp. 6 and 15; result and pretraining discussion p. 7 · Main paper
What to reuse—and what not to infer
Transferable lesson: For multimodal alignment, draw training signals separately from retrieval outputs, define each generalization axis, and align every result block and ablation with those axes.
Boundary: Zero-shot here concerns retrieval for held-out concepts under defined candidate sets; it is not unrestricted mind reading. Temporal/spatial/spectral associations do not establish causal neuroscience, and the author version is not verified camera-ready text.
Apply this lesson using the argument and evidence worksheet.
Inspection record and sources
- Checked
- 2026-10-02
- Reading status
- Writing-focused manuscript scan + selected visuals inspected
- Main text
- Writing-oriented scan across manuscript framing, section sequence, figure/table captions, selected argument/evaluation prose and conclusion; not exhaustive technical or proof verification. Sections 1–5, pp. 1–10; references pp. 10–14. Rendered the tri-modal alignment/retrieval overview p. 2 and subject-wise evidence plus ablations p. 7. Source is the exact public author version listed below, not a verified accepted/camera-ready revision.
- Appendix / supplement
- Writing-oriented appendix/documentation scan: Appendices A–D, pp. 15–20: data and preprocessing, implementation, per-subject/encoder/prompt comparisons, qualitative retrieval and additional visualizations. Source code, data, model outputs and external video assets were not executed or independently validated.
- Visually inspected PDF pages
- 2, 7
This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.