Writing study: Seeing Through the Brain: New Insights from Decoding Visual Stimuli with fMRI

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Zheng Huang, Enpei Zhang, Weikang Qiu, Yinghao Cai, Carl Yang, Elynn Chen, Xiang Zhang, Rex Ying, Dawei Zhou, Yujun Yan · Official ICLR 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Architecture & representation design (primary) · Empirical analysis, diagnosis & auditing.

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

25 physical pages

Before references / appendix

≈ 10.32 page-equivalents

Main figures / tables

2 / 6

Attached appendix figures / tables

5 / 3

Attached appendix span

≈ 8.89 page-equivalents

Before references / appendix: ≈ 41.28%References / intervening material: ≈ 23.16%Attached appendix: ≈ 35.56%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

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	≈ 0.97 pp.	3.88%
2 PRELIMINARY	2	≈ 0.96 pp.	3.84%
3 METHOD	3	≈ 2.97 pp.	11.88%
4 EXPERIMENTS	6	≈ 3.16 pp.	12.64%
5 RELATED WORK	9	≈ 1 pp.	4%
6 CONCLUSION	10	≈ 0.38 pp.	1.52%
7 ETHICS STATEMENT	11	≈ 0.11 pp.	0.44%
8 REPRODUCIBILITY STATEMENT	11	≈ 0.11 pp.	0.44%
REFERENCES	11	≈ 5.78 pp.	23.12%
A A COMMON ERROR IN GENERATIVE MODELS	17	≈ 0.49 pp.	1.96%
B PRELIMINARY: DIFFUSION MODEL	17	≈ 0.81 pp.	3.24%
C DATASET	18	≈ 0.7 pp.	2.8%
D BASELINES	19	≈ 0.63 pp.	2.52%
E TRAINING DETAILS	19	≈ 0.59 pp.	2.36%
F CASE STUDY ON OBJECT-LEVEL DESCRIPTIONS FOR IMAGE	20	≈ 0.47 pp.	1.88%
G RECONSTRUCTION PERFORMANCE ACROSS DIFFERENT LATENT SPACES	20	≈ 0.46 pp.	1.84%
H ANALYSIS ON THE NUMBER OF OBJECTS IN OUR FRAMEWORK	21	≈ 1.15 pp.	4.6%
I ATTRIBUTE/RELATIONSHIP OPTIMIZATION DETAILS	22	≈ 0.69 pp.	2.76%
J KEYWORD SEARCH WITH DIFFERENT INITIAL SETS	23	≈ 0.26 pp.	1.04%
K PROMPTS USED BY PRISM	23	≈ 1.74 pp.	6.96%
L THE USE OF LARGE LANGUAGE MODELS (LLMS)	25	≈ 0.89 pp.	3.56%

All detected figure and table locations

Main text

• Figure 1 · p. 3
• Table 1 · p. 4
• Table 2 · p. 7
• Table 3 · p. 7
• Figure 2 · p. 8
• Table 4 · p. 8
• Table 5 · p. 8
• Table 6 · p. 10

Attached appendix

• Figure 3 · p. 17
• Table 7 · p. 20
• Figure 4 · p. 20
• Table 8 · p. 21
• Figure 5 · p. 21
• Figure 6 · p. 22
• Table 9 · p. 23
• Figure 7 · p. 24

Argument and organization

1. Representation comparison

   Motivates structured text as the intermediate space before proposing PRISM.

2. Description search, encoder training and object composition

   Separates training-time image annotation from fMRI-only reconstruction.

3. Research-question experiments

   Tests reconstruction, latent-space choice and module contributions in order.

Problem, claim and contribution

Rather than opening with a new backbone, the paper asks which intermediate representation should connect brain signals and an image generator. The representation comparison supplies a rationale for the method.

Sections 1 and 3.1, pages 1–4 · Main paper

Argument and section order

Three explicit research questions map to the main performance table and ablations, while the keyword-search case study tests a more specific interpretation of the method.

Sections 4.1–4.4, pages 6–9 · Main paper

Figures, captions and visual explanation

Figure 1 separates annotation/search, supervised fMRI-to-text training and inference with horizontal bands. Its repeated elephant-and-car example makes the symbolic object descriptions concrete.

Figure 1, page 3 · Main paper

Tables, comparisons and evidence

Table 2 groups results by dataset and separates SDXL variants from the shared SD 2.1 backbone, avoiding a misleading comparison across different generators.

Table 2, page 7; Appendix D, page 19 · Main paper

Appendix and reproducibility

Exact description templates and an epsilon-greedy search algorithm expose otherwise hidden choices: two object slots, discrete spatial labels and a 667-image search subset.

Appendices H–K, pages 21–24 · Main paper

Clarity and writing conventions

Brain-alignment claims should remain empirical and model-dependent. The paper appropriately notes that its ROI interpretation needs further neuroscience validation, but some broader conclusions use stronger language than those comparisons warrant.

Section 4.4, page 9; conclusion page 10 · Main paper

What to reuse—and what not to infer

Transferable lesson: Use a motivating representation comparison to justify architecture, then make training-only supervision visibly distinct from information available at inference.

Boundary: Higher similarity to selected text encoders is not a general proof that the brain represents images as language. The practical model fixes two object groups and coarse spatial labels; its generated images are reconstructions, not direct visual readouts.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

Pages 1–10, Sections 1–6

Appendix / supplement

Pages 17–25, Appendices A–L, including prompt templates, search algorithm and training settings

Visually inspected PDF pages

3, 7

• ICLR 2026 final proceedings · 25 pages