Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Escaping the BLEU Trap: A Signal-Grounded Framework with Decoupled Semantic Guidance for EEG-to-Text Decoding

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SemKey, a new EEG-to-text framework, forces a language model to ground every generated word in neural signals by first predicting sentiment, topic, length, and surprisal from brainwave embeddings and then routing those predictions into a qu

desk verdict A useful evaluation critique and a plausible two-stage framework, but the signal-grounded claim is unproven—attribute heads may exploit source corpus confounds, and content recall is 2.7%. read the letter →

arxiv 2603.03312 v3 pith:I2CIDZGP submitted 2026-02-09 cs.CL cs.AIcs.HCeess.ASq-bio.NC

Yuchen Wang , Haonan Wang , Yu Guo , Honglong Yang , Xiaomeng Li This is my paper
classification cs.CLcs.AIcs.HCeess.ASq-bio.NC
keywords EEG-to-textdecodingsignalgroundingsemanticattributesQ-K-VinjectionBLEUtraphallucinationretrievalaccuracybrain-computerinterface
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that existing EEG-to-text decoders are not really decoding: they rely on teacher forcing at test time, collapse into generic templates like "He was...", and hallucinate coherent sentences even when fed pure Gaussian noise. To fix this, SemKey first extracts four high-level semantic attributes—sentiment, topic, length, and surprisal—from EEG embeddings, then uses a Q-K-V injection workflow so the frozen language model must attend to neural signals rather than linguistic priors. The authors also contend that standard BLEU scores are misleading because they reward stopword-heavy template matching; they propose retrieval-based accuracy, content recall, diversity metrics, and Fréchet distance as a more honest evaluation protocol. If the framework works as claimed, it would give the field a way to build and verify genuinely signal-grounded brain-to-text systems for non-invasive BCIs.

What carries the argument

The central mechanism is the Q-K-V Injection workflow: the autoregressive text prompt and generated tokens are processed to form queries, while a concatenation of the global EEG vector and token-level EEG embeddings is projected into keys and values. This reverses the usual conditioning setup so the LLM must actively retrieve information from neural signals at every decoding step. Complementing this is the Parallel Neural-Driven Attribute Decouple, four lightweight prediction heads (sentiment classifier, topic classifier, length regressor, surprisal regressor) that convert EEG features into explicit semantic constraints, which are then verbalized into a structured prompt that steers generati

What would settle it

Take the trained SemKey model and replace the Stage-1 attribute predictions with random labels (or the opposite sentiment/topic) while keeping the EEG embeddings intact; if generation quality and retrieval accuracy do not drop substantially, the claimed signal-grounded steering is not actually attributable to the decoded attributes. A second decisive test is to train a variant without the Q-K-V injection but with the same attribute prompt; if that variant still passes the Gaussian-noise test (i.e., produces gibberish) and matches SemKey's retrieval scores, then the Q-K-V mechanism is not the d

Watch

Extended reading notes

Core claim

SemKey replaces the conventional end-to-end EEG-to-text pipeline with a two-stage 'guidance-generation' paradigm. In Stage 1, a Conformer-based EEG encoder is trained with alignment and reconstruction losses plus four auxiliary heads that predict sentiment, topic, length, and surprisal, forming a structured semantic anchor. In Stage 2, these predicted attributes are folded into a natural-language prompt, and the prompt's hidden states become queries while projected EEG embeddings and a global EEG vector become keys and values in the LLM's cross-attention. This forces token-by-token generation to depend on the input signal. The paper shows that when real EEG is replaced with Gaussian noise, S

Load-bearing premise

The four semantic attributes (sentiment, topic, length, surprisal) must be decodable from EEG accurately enough to steer generation, and they must be sufficient as generative constraints; if they are not, the prompt could be injecting dataset priors rather than genuine neural semantics.

Editorial extensions

If this is right

  • If SemKey's signal-grounding holds, EEG-to-text systems can be evaluated without teacher forcing, making them suitable for real-world BCI use where the user's intended message is unknown.
  • The noise-input test becomes a practical sanity check: any EEG-to-text model that produces fluent text from Gaussian noise is likely relying on language priors rather than decoding neural signals.
  • The proposed evaluation protocol—N-way retrieval accuracy, content recall, head entropy, self-BLEU, and Fréchet distance—could become a standard for gauging semantic fidelity and diversity in brain-to-text research, revealing template collapse that BLEU hides.
  • The finding that full LLM fine-tuning or LoRA hurts performance on small EEG datasets suggests that frozen LLMs plus trainable projections is the right regime for low-resource neural decoding.
  • Removing any single semantic attribute degrades performance, indicating that the four anchors are complementary and jointly constrain the generative search space.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same attribute-decoupling plus Q-K-V injection recipe could generalize to other low-resource neural decoding tasks (e.g., fMRI-to-text or ECoG-to-speech), where hallucination from strong priors is equally problematic.
  • The four chosen attributes are only a first slice of semantic content; richer labels such as named-entity presence, temporal structure, or discourse role might push fidelity further, especially for proper nouns which the paper admits remain hard.
  • The 'BLEU Trap' argument likely applies beyond EEG-to-text—any conditional generation task with narrow references and template-prone models can inflate n-gram metrics, so the proposed evaluation set could transfer to summarization or dialogue.
  • A standardized 'signal-destruction curve'—progressively mixing in Gaussian noise and measuring content recall and diversity—could become a reproducibility benchmark for grounding claims across brain-decoding papers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes SemKey, a two-stage EEG-to-text decoding framework that aims to address three failure modes: semantic bias (template collapse), signal neglect (hallucination from LLM priors), and the 'BLEU Trap' (n-gram metrics inflated by stopwords/templates). Stage 1 trains an EEG encoder with four auxiliary heads that predict sentiment, topic, length, and surprisal from EEG embeddings. Stage 2 freezes the attribute heads and injects the predicted attributes as a text prompt into a frozen Flan-T5 decoder, while rerouting cross-attention so the prompt tokens act as Queries and projected EEG embeddings as Keys/Values. The paper also introduces a new evaluation protocol (N-way retrieval accuracy with Monte Carlo sampling, content recall, self-BLEU, head entropy, and Fréchet Distance) and reports state-of-the-art results over the GLIM baseline on ZuCo 1.0/2.0, including a noise-input test in which SemKey degrades to gibberish while GLIM continues to produce fluent but hallucinated text.

Significance. If the central 'signal-grounded' claim were fully established, this would be a substantively important contribution: it would demonstrate a practical path away from teacher-forced, template-collapsing EEG-to-text models and provide a more robust evaluation protocol for the field. The paper has notable strengths: a comprehensive ablation study (Tables 5, b, h), statistical reporting over 7 seeds (Appendix A), a noise-degradation curve (Appendix D), and an unusually candid limitations appendix (Appendix I) that acknowledges content recall of 2.7%, 4-way retrieval near 50%, and prominent factual errors in qualitative outputs. The proposed evaluation metrics, especially Monte Carlo N-way retrieval and the noise-collapse test, are valuable methodological additions regardless of the specific model. However, the paper's headline claim that SemKey 'enforces signal-grounded generation' and achieves 'semantic fidelity' is not yet supported by the evidence, because the attribute heads may be exploiting dataset/source confounds rather than decoding per-trial neural semantics, and the new metrics are introduced and optimized by the same method without external validation. The contribution is th

major comments (3)
  1. [§3.1, Table 4] The signal-grounded claim rests on the four attribute heads, but Table 4 is not sufficient to rule out confounds. The topic labels are exactly the two ZuCo source corpora ('Biographies and Factual Knowledge' vs 'Movie Reviews and Sentiment'), the sentiment labels are binary, and the EEG encoder is conditioned on dataset/subject priors via the context vector described in §3.1. Topic accuracy of 83.5% and sentiment accuracy of 90.7% could therefore reflect corpus/reading-task separability rather than per-trial EEG semantic decoding. Length (MAE 5.01 vs chance 6.23) and surprisal (MAE 0.58 vs chance 0.78) are only modestly above chance. If the attribute heads are mostly detecting source corpus, then the prompt in §3.2 essentially tells the frozen LLM which genre to produce, which could explain the diversity gains in Table 5 without demonstrating neural semantic grounding. Please provide con
  2. [§4.2, Table 3] The noise-input test shows that SemKey is input-dependent, but it does not establish that real EEG is the source of semantic content. Any cross-attention decoder fed out-of-distribution Gaussian K/V vectors can collapse to disordered tokens. The paper's interpretation that this 'strictly prevents the generation of coherent text in the absence of valid neural information' conflates sensitivity to input distribution with neural decoding fidelity. A more informative control would use real but uninformative neural inputs, e.g., phase-scrambled EEG, EEG recorded during a different task, or trial-to-text labels shuffled. If SemKey also collapses under those conditions, the result would be less specific to signal absence; if it does not, that would directly support the signal-grounded claim. The current test is a useful sanity check but not proof of semantic grounding.
  3. [§3.3, Table 1, Appendix I, Table m] The paper introduces new evaluation metrics (N-way retrieval, content recall, FD) and then reports SOTA on them, while several model components explicitly optimize diversity and distributional similarity. This creates a mild circularity that should be addressed by validating the metrics against human judgments or an external benchmark. The qualitative evidence in Table m and the authors' own limitations section (Appendix I: content recall 2.7%, 4-way accuracy 50%) show frequent factual errors (e.g., 'Libby Carrie Fisher (born June 9, 1916) is an American...' for a ground truth about Taylor; 'During the following year, she married Michael Jackson...' for a ground truth about Chynna Phillips). The retrieval/FD improvements do not yet demonstrate 'semantic fidelity' in the sense the abstract claims. Please include a human evaluation of factuality or at least a quantitative analysis of named
minor comments (5)
  1. [§3.2 vs Appendix C.2] There is a contradiction about whether the LLM is frozen. §3.2 states 'the LLM Decoder is frozen,' but Appendix C.2 says 'the learning rates were layer-specific: 8.0×10−5 for the Stage 1 encoder, 2.0×10−4 for the projector, and 1.0×10−5 for the unfrozen LLM components.' Please clarify which components are actually trainable in Stage 2.
  2. [Abstract vs Checklist] The abstract says 'Code will be released upon acceptance,' while the NeurIPS checklist says 'Our code in attached in supplementary materials.' If code is attached as supplementary material, please state that explicitly; if not, the checklist response is inaccurate.
  3. [Appendix A] The t-tests are reported as p < 0.001 for all metrics without correction for multiple comparisons. Since eight metrics are tested, please report adjusted p-values or note the number of comparisons and whether the conclusions survive correction.
  4. [Eq. (4), Table 1] The Fréchet Distance is computed on sentence embeddings from a pre-trained Sentence Transformer, but Eq. (4) assumes the embeddings are approximately Gaussian. With 2227 test sentences and high-dimensional embeddings, the covariance estimate may be unstable. Please comment on this or consider a non-parametric alternative (e.g., sliced-Wasserstein distance or nearest-neighbor based divergence).
  5. [Table 2] The phrase 'MTV Evaluation' is used for augmenting references with paraphrases, but the term 'MTV' is not expanded in the main text (it appears in Appendix B as 'Multiple Text Variants'). Please define it at first use in the main text.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: main results are empirical comparisons on externally applicable metrics; self-citations are not load-bearing.

full rationale

SemKey's derivation chain is a supervised attribute-prediction stage followed by a prompt/Q-K-V injection into a frozen LLM and a new evaluation protocol. None of these steps reduces to its own inputs by construction: the attribute heads are trained on ZuCo text-derived labels from EEG features, the Q-K-V workflow is a concrete cross-attention variant (Eq. 2), and the retrieval/diversity/FD metrics are defined independently of SemKey and applied to all baselines with the same protocol. The reported SOTA is therefore an empirical comparison, not a tautology. Self-citations [29], [39], and [40] are used for motivation/design choices (parallel multi-task, inter-subject variability) rather than as the sole support for the central claim; the paper's own ablations and noise tests provide independent evidence for input dependence. The topic/sentiment labels are coarse and the encoder is conditioned on dataset/subject priors, which raises a validity concern about whether 'semantic anchors' are driven by EEG or source confounds; however, the paper does not define topic as a function of the prior, so this is a threat to the interpretation of the results, not a circular reduction. Under the evidence quoted, no step is equivalent to its input by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; all components are configurations of existing models (Conformer, Flan-T5) and standard text attributes. The main load-bearing assumptions are the decodability/sufficiency of the four attributes and the interpretation of noise-collapse as signal grounding.

free parameters (3)
  • Loss weights λ = Stage1: L_align=0.5, L_recon=0.5, L_cls=0.3, L_reg=0.3-0.9; E2E: L_recon=1.5, L_cls=0.25, L_reg=0.25
    Hand-selected and tuned on validation; these weights balance alignment, reconstruction, and auxiliary objectives, directly controlling guidance strength.
  • Semantic attribute set = sentiment / topic / length / surprisal (2-class topic, 2-class sentiment)
    Ad hoc design choice; no evidence presented that these four attributes are sufficient or optimal, and the label spaces are coarse.
  • Monte Carlo trials M = 1000
    Number of sampling trials in N-way retrieval accuracy; chosen to reduce variance, not derived from a principled error target.
assumptions (4)
  • domain assumption EEG signals contain decodable high-level semantic information corresponding to sentiment, topic, length, and surprisal.
    Section 3.1 invokes cortical representation literature [21], but the paper's own evidence (Table 4) shows only moderate predictive accuracy.
  • domain assumption ZuCo 1.0/2.0 sentence-level EEG-text pairs are a valid and representative benchmark for non-invasive EEG-to-text decoding.
    Section 4 and Appendix C.3; used by prior SOTA, but limited to reading-task EEG and may not generalize beyond this paradigm.
  • domain assumption The Q-K-V injection forces the frozen LLM to base generation on EEG-derived keys/values; noise collapse to gibberish is evidence of signal grounding, not a degenerate failure.
    Section 3.2 and Section 4.2; alternative explanation: out-of-distribution noise disrupts attention and causes collapse, which would not prove signal grounding.
  • standard math Fréchet distance computed on sentence-transformer embeddings is a valid distribution-alignment metric under the Gaussian assumption.
    Equation 4; assumes embeddings are multivariate Gaussians, which is not verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Escaping the BLEU Trap: A Signal-Grounded Framework with Decoupled Semantic Guidance for EEG-to-Text Decoding." pith.science (2026). https://pith.science/paper/I2CIDZGP

@misc{pith2026260303312,
  author       = {Pith},
  title        = {Pith review of: Escaping the BLEU Trap: A Signal-Grounded Framework with Decoupled Semantic Guidance for EEG-to-Text Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I2CIDZGP}},
  note         = {Machine review of arXiv:2603.03312}
}
read the original abstract

Decoding natural language from non-invasive EEG signals is a promising yet challenging task. However, current state-of-the-art models remain constrained by three fundamental issues: Semantic Bias, where outputs collapse into generic linguistic templates; Signal Neglect, where models rely heavily on LLM priors to hallucinate fluent text even in the absence of meaningful signals; and the "BLEU Trap", where high-frequency stopwords inflate n-gram metrics, masking a lack of true semantic fidelity. To resolve these challenges, we move beyond conventional end-to-end pipelines and propose SemKey, a novel multi-stage framework that enforces signal-grounded generation through four decoupled semantic objectives: sentiment, topic, length, and surprisal. We extract these semantic anchors from EEG embeddings directly, then unify them with an Active Retrieval Decoding mechanism, compelling the LLM to ground its token generation in the neural signals rather than defaulting to linguistic priors. Furthermore, we break the BLEU Trap by establishing a comprehensive evaluation protocol using rigorous retrieval and distribution-based metrics such as Fr\'echet Distance. Extensive experiments demonstrate that SemKey effectively mitigates hallucinations on noise inputs and achieves SOTA performance on these robust protocols. Code will be released upon acceptance at https://github.com/xmed-lab/SemKey.

Figures

Figures reproduced from arXiv: 2603.03312 by the authors.

Figure 1
Figure 1. Left (Decoding showcase & Quantitative Results): Previous models [27] exhibit se￾vere semantic bias, overfitting to generic templates to artificially inflate BLEU scores despite low semantic alignment and repetition - a phenomenon we identify as the “BLEU Trap”. The noise test (gray highlight) exposes signal neglect: the model continues to hallucinate fluent text from pure Gaussian noise. In contrast, SEMKEY breaks … view at source ↗
Figure 2
Figure 2. The overall architecture of the SEMKEY framework following a "Guidance-Generation" paradigm. Stage 1 (Attribute Extraction): The EEG encoder is optimized via multi-task learning to explicitly decouple high-level semantic attributes (Sentiment, Topic, Length, Surprisal) alongside standard alignment. Stage 2 (Generative Decoding): These predicted attributes structure a semantic prompt to guide generation. Crucially, t… view at source ↗
Figure 3
Figure 3. Holistic Performance Evaluation. The Radar Chart illustrates the relative performance of our model (Or￾ange) versus GLIM (Blue) and prior model without teacher forcing (Green). SEMKEY demonstrates a balanced supe￾riority across all alignment, diversity, and content quality metrics. We conduct our experiments using the data from the ZuCo 1.0 [9] and ZuCo 2.0 [10] datasets, containing over 22K sentence-EEG pairs (deta… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Showcase of Gaussian Noise input. We illustrate the text generated by GLIM versus our method when the input is pure Gaussian noise instead of EEG signals. To verify the signal dependency of our frame￾work, we conducted a Noise Input Test, where real EEG signals were re…
Figure 5
Figure 5. Figure 5: t-SNE Visualization of Semantic Distribution. We visualize the sentence embeddings of Ground Truth (Blue), SEMKEY (Orange), GLIM (Yellow), and Prior Model(w/o tf) (Purple). Our model shows better overlap with the ground truth distribution, while effectively avoiding se…

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 16 linked inside Pith

  1. [1]

    Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35:23716–23736, 2022

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35:23716–23736, 2022

    2022

  2. [2]

    Scaling instruction-finetuned language models

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70):1–53, 2024

    2024

  3. [3]

    Mel frequency cepstral coefficients enhance imagined speech decoding accuracy from eeg

    Ciaran Cooney, Rafaella Folli, and Damien Coyle. Mel frequency cepstral coefficients enhance imagined speech decoding accuracy from eeg. In2018 29th Irish Signals and Systems Conference (ISSC), pages 1–7. IEEE, 2018

    2018

  4. [4]

    Decoding speech perception from non-invasive brain recordings.Nature Machine Intelligence, 5(10):1097–1107, 2023

    Alexandre Défossez, Charlotte Caucheteux, Jérémy Rapin, Ori Kabeli, and Jean-Rémi King. Decoding speech perception from non-invasive brain recordings.Nature Machine Intelligence, 5(10):1097–1107, 2023

    2023

  5. [5]

    Dewave: Discrete eeg waves encoding for brain dynamics to text translation.arXiv preprint arXiv:2309.14030, 2023

    Yiqun Duan, Jinzhao Zhou, Zhen Wang, Yu-Kai Wang, and Chin-Teng Lin. Dewave: Discrete eeg waves encoding for brain dynamics to text translation.arXiv preprint arXiv:2309.14030, 2023

    arXiv 2023

  6. [6]

    Acoustic inspired brain-to-sentence decoder for logosyllabic language.Cyborg and Bionic Systems, 6:0257, 2025

    Chen Feng, Lu Cao, Di Wu, En Zhang, Ting Wang, Xiaowei Jiang, Jinbo Chen, Hui Wu, Siyu Lin, Qiming Hou, et al. Acoustic inspired brain-to-sentence decoder for logosyllabic language.Cyborg and Bionic Systems, 6:0257, 2025

    2025

  7. [7]

    Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity.Nature neuroscience, 18(11):1664–1671, 2015

    Emily S Finn, Xilin Shen, Dustin Scheinost, Monica D Rosenberg, Jessica Huang, Marvin M Chun, Xenophon Papademetris, and R Todd Constable. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity.Nature neuroscience, 18(11):1664–1671, 2015

    2015

  8. [8]

    Conformer: Convolution-augmented transformer for speech recognition.arXiv preprint arXiv:2005.08100, 2020

    Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer for speech recognition.arXiv preprint arXiv:2005.08100, 2020

    arXiv 2005

Show all 52 references
  1. [9]

    Zuco, a simultaneous eeg and eye-tracking resource for natural sentence reading.Scientific data, 5(1): 1–13, 2018

    Nora Hollenstein, Jonathan Rotsztejn, Marius Troendle, Andreas Pedroni, Ce Zhang, and Nicolas Langer. Zuco, a simultaneous eeg and eye-tracking resource for natural sentence reading.Scientific data, 5(1): 1–13, 2018

    2018

  2. [10]

    Zuco 2.0: A dataset of physiological recordings during natural reading and annotation

    Nora Hollenstein, Marius Troendle, Ce Zhang, and Nicolas Langer. Zuco 2.0: A dataset of physiological recordings during natural reading and annotation. InProceedings of the Twelfth Language Resources and Evaluation Conference, pages 138–146, 2020

    2020

  3. [11]

    The curious case of neural text degeneration.arXiv preprint arXiv:1904.09751, 2019

    Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration.arXiv preprint arXiv:1904.09751, 2019

    1904 arXiv

  4. [12]

    Natural speech reveals the semantic maps that tile human cerebral cortex.Nature, 532(7600):453–458, 2016

    Alexander G Huth, Wendy A De Heer, Thomas L Griffiths, Frédéric E Theunissen, and Jack L Gallant. Natural speech reveals the semantic maps that tile human cerebral cortex.Nature, 532(7600):453–458, 2016

    2016

  5. [13]

    Survey of hallucination in natural language generation.ACM computing surveys, 55(12):1–38, 2023

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation.ACM computing surveys, 55(12):1–38, 2023

    2023

  6. [14]

    Large brain model for learning generic representations with tremendous eeg data in bci.arXiv preprint arXiv:2405.18765, 2024

    Wei-Bang Jiang, Li-Ming Zhao, and Bao-Liang Lu. Large brain model for learning generic representations with tremendous eeg data in bci.arXiv preprint arXiv:2405.18765, 2024. 10

    2024 arXiv

  7. [15]

    Are eeg-to-text models working?arXiv preprint arXiv:2405.06459, 2024

    Hyejeong Jo, Yiqian Yang, Juhyeok Han, Yiqun Duan, Hui Xiong, and Won Hee Lee. Are eeg-to-text models working?arXiv preprint arXiv:2405.06459, 2024

    2024 arXiv

  8. [16]

    A capacity theory of comprehension: individual differences in working memory.Psychological review, 99(1):122, 1992

    Marcel A Just and Patricia A Carpenter. A capacity theory of comprehension: individual differences in working memory.Psychological review, 99(1):122, 1992

    1992

  9. [17]

    Ctrl: A conditional transformer language model for controllable generation.arXiv preprint arXiv:1909.05858, 2019

    Nitish Shirish Keskar, Bryan McCann, Lav R Varshney, Caiming Xiong, and Richard Socher. Ctrl: A conditional transformer language model for controllable generation.arXiv preprint arXiv:1909.05858, 2019

    1909 arXiv

  10. [18]

    Bendr: Using transformers and a contrastive self-supervised learning task to learn from massive amounts of eeg data.Frontiers in Human Neuroscience, 15:653659, 2021

    Demetres Kostas, Stephane Aroca-Ouellette, and Frank Rudzicz. Bendr: Using transformers and a contrastive self-supervised learning task to learn from massive amounts of eeg data.Frontiers in Human Neuroscience, 15:653659, 2021

    2021

  11. [19]

    Reading senseless sentences: Brain potentials reflect semantic incongruity.Science, 207(4427):203–205, 1980

    Marta Kutas and Steven A Hillyard. Reading senseless sentences: Brain potentials reflect semantic incongruity.Science, 207(4427):203–205, 1980

    1980

  12. [20]

    Set transformer: A framework for attention-based permutation-invariant neural networks

    Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. InInternational conference on machine learning, pages 3744–3753. PMLR, 2019

    2019

  13. [21]

    Topographic mapping of a hierarchy of temporal receptive windows using a narrated story.Journal of neuroscience, 31(8):2906–2915, 2011

    Yulia Lerner, Christopher J Honey, Lauren J Silbert, and Uri Hasson. Topographic mapping of a hierarchy of temporal receptive windows using a narrated story.Journal of neuroscience, 31(8):2906–2915, 2011

    2011

  14. [22]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021

    2021 arXiv

  15. [23]

    Visual decoding and reconstruction via eeg embeddings with guided diffusion.arXiv preprint arXiv:2403.07721, 2024

    Dongyang Li, Chen Wei, Shiying Li, Jiachen Zou, Haoyang Qin, and Quanying Liu. Visual decoding and reconstruction via eeg embeddings with guided diffusion.arXiv preprint arXiv:2403.07721, 2024

    2024 arXiv

  16. [24]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  17. [25]

    Controllable text generation for large language models: A survey.arXiv preprint arXiv:2408.12599, 2024

    Xun Liang, Hanyu Wang, Yezhaohui Wang, Shichao Song, Jiawei Yang, Simin Niu, Jie Hu, Dan Liu, Shunyu Yao, Feiyu Xiong, et al. Controllable text generation for large language models: A survey.arXiv preprint arXiv:2408.12599, 2024

    2024 arXiv

  18. [26]

    Eeg2text: Open vocabulary eeg-to-text translation with multi-view transformer

    Hanwen Liu, Daniel Hajialigol, Benny Antony, Aiguo Han, and Xuan Wang. Eeg2text: Open vocabulary eeg-to-text translation with multi-view transformer. In2024 IEEE International Conference on Big Data (BigData), pages 1824–1833. IEEE, 2024

    2024

  19. [27]

    Learning interpretable representations leads to semantically faithful eeg-to-text generation.arXiv preprint arXiv:2505.17099, 2025

    Xiaozhao Liu, Dinggang Shen, and Xihui Liu. Learning interpretable representations leads to semantically faithful eeg-to-text generation.arXiv preprint arXiv:2505.17099, 2025

    2025 arXiv

  20. [28]

    How to control sentiment in text generation: A survey of the state-of-the- art in sentiment-control techniques

    Michela Lorandi and Anja Belz. How to control sentiment in text generation: A survey of the state-of-the- art in sentiment-control techniques. InProceedings of the 13th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis, pages 341–353, 2023

    2023

  21. [29]

    A cognitive process-inspired architecture for subject-agnostic brain visual decoding

    Jingyu Lu, Haonan Wang, Qixiang Zhang, and Xiaomeng Li. A cognitive process-inspired architecture for subject-agnostic brain visual decoding. InThe Fourteenth International Conference on Learning Representations (ICLR), 2026

    2026

  22. [30]

    On faithfulness and factuality in abstractive summarization.arXiv preprint arXiv:2005.00661, 2020

    Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. On faithfulness and factuality in abstractive summarization.arXiv preprint arXiv:2005.00661, 2020

    2005 arXiv

  23. [31]

    A high-performance neuroprosthesis for speech decoding and avatar control.Nature, 620(7976):1037–1046, 2023

    Sean L Metzger, Kaylo T Littlejohn, Alexander B Silva, David A Moses, Margaret P Seaton, Ran Wang, Maximilian E Dougherty, Jessie R Liu, Peter Wu, Michael A Berger, et al. A high-performance neuroprosthesis for speech decoding and avatar control.Nature, 620(7976):1037–1046, 2023

    2023

  24. [32]

    Strong prediction: Language model surprisal explains multiple n400 effects.Neurobiology of language, 5 (1):107–135, 2024

    James A Michaelov, Megan D Bardolph, Cyma K Van Petten, Benjamin K Bergen, and Seana Coulson. Strong prediction: Language model surprisal explains multiple n400 effects.Neurobiology of language, 5 (1):107–135, 2024

    2024

  25. [33]

    Thought2text: text generation from eeg signal using large language models (llms)

    Abhijit Mishra, Shreya Shukla, Jose Torres, Jacek Gwizdka, and Shounak Roychowdhury. Thought2text: text generation from eeg signal using large language models (llms). InFindings of the Association for Computational Linguistics: NAACL 2025, pages 3747–3759, 2025. 11

    2025

  26. [34]

    A novel deep learning architecture for decoding imagined speech from eeg.arXiv preprint arXiv:2003.09374, 2020

    Jerrin Thomas Panachakel, AG Ramakrishnan, and TV Ananthapadmanabha. A novel deep learning architecture for decoding imagined speech from eeg.arXiv preprint arXiv:2003.09374, 2020

    2003 arXiv

  27. [35]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

    2021

  28. [36]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019

    1908 arXiv

  29. [37]

    On accurate evaluation of gans for language generation.arXiv preprint arXiv:1806.04936, 2018

    Stanislau Semeniuta, Aliaksei Severyn, and Sylvain Gelly. On accurate evaluation of gans for language generation.arXiv preprint arXiv:1806.04936, 2018

    2018 arXiv

  30. [38]

    See: Semantically aligned eeg-to-text translation

    Yitian Tao, Yan Liang, Luoyu Wang, Yongqing Li, Qing Yang, and Han Zhang. See: Semantically aligned eeg-to-text translation. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025

    2025

  31. [39]

    Zebra: Towards zero-shot cross-subject generalization for universal brain visual decoding.arXiv preprint arXiv:2510.27128, 2025

    Haonan Wang, Jingyu Lu, Hongrui Li, and Xiaomeng Li. Zebra: Towards zero-shot cross-subject generalization for universal brain visual decoding.arXiv preprint arXiv:2510.27128, 2025

    2025

  32. [40]

    Neurons: Emulating the human visual cortex improves fidelity and interpretability in fmri-to-video reconstruction.arXiv preprint arXiv:2503.11167, 2025

    Haonan Wang, Qixiang Zhang, Lehan Wang, Xuanqi Huang, and Xiaomeng Li. Neurons: Emulating the human visual cortex improves fidelity and interpretability in fmri-to-video reconstruction.arXiv preprint arXiv:2503.11167, 2025

    2025 arXiv

  33. [41]

    Open vocabulary electroencephalography-to-text decoding and zero-shot sentiment classification

    Zhenhailong Wang and Heng Ji. Open vocabulary electroencephalography-to-text decoding and zero-shot sentiment classification. InProceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 5350–5358, 2022

    2022

  34. [42]

    A high-performance speech neuroprosthesis.Nature, 620(7976):1031–1036, 2023

    Francis R Willett, Erin M Kunz, Chaofei Fan, Donald T Avansino, Guy H Wilson, Eun Young Choi, Foram Kamdar, Matthew F Glasser, Leigh R Hochberg, Shaul Druckmann, et al. A high-performance speech neuroprosthesis.Nature, 620(7976):1031–1036, 2023

    2023

  35. [43]

    Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks

    Nan Wu, Stanislaw Jastrzebski, Kyunghyun Cho, and Krzysztof J Geras. Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks. InInternational Conference on Machine Learning, pages 24043–24055. PMLR, 2022

    2022

  36. [44]

    Wavemind: Towards a conversational eeg foundation model aligned to textual and visual modalities.arXiv preprint arXiv:2510.00032, 2025

    Ziyi Zeng, Zhenyang Cai, Yixi Cai, Xidong Wang, Junying Chen, Rongsheng Wang, Yipeng Liu, Siqi Cai, Benyou Wang, Zhiguo Zhang, et al. Wavemind: Towards a conversational eeg foundation model aligned to textual and visual modalities.arXiv preprint arXiv:2510.00032, 2025

    2025

  37. [45]

    Converting your thoughts to texts: Enabling brain typing via deep feature learning of eeg signals

    Xiang Zhang, Lina Yao, Quan Z Sheng, Salil S Kanhere, Tao Gu, and Dalin Zhang. Converting your thoughts to texts: Enabling brain typing via deep feature learning of eeg signals. In2018 IEEE international conference on pervasive computing and communications (PerCom), pages 1–10...

    2018

  38. [46]

    Du-in: Discrete units-guided mask modeling for decoding speech from intracranial neural signals.Advances in Neural Information Processing Systems, 37:79996–80033, 2024

    Hui Zheng, Haiteng Wang, Weibang Jiang, Zhongtao Chen, Li He, Peiyang Lin, Penghu Wei, Guoguang Zhao, and Yunzhe Liu. Du-in: Discrete units-guided mask modeling for decoding speech from intracranial neural signals.Advances in Neural Information Processing Systems, 37:79996–80033, 2024

    2024

  39. [47]

    Pretraining large brain language model for active bci: Silent speech.arXiv preprint arXiv:2504.21214, 2025

    Jinzhao Zhou, Zehong Cao, Yiqun Duan, Connor Barkley, Daniel Leong, Xiaowei Jiang, Quoc-Toan Nguyen, Ziyi Zhao, Thomas Do, Yu-Cheng Chang, et al. Pretraining large brain language model for active bci: Silent speech.arXiv preprint arXiv:2504.21214, 2025

    2025 arXiv

  40. [48]

    BLEU Trap

    Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu. Texygen: A benchmarking platform for text generation models. InThe 41st international ACM SIGIR conference on research & development in information retrieval, pages 1097–1100, 2018. 12 Supplemen...

    2018

  41. [49]

    LLM Tuning Strategy LLM FULLFINE-TUNE47.9% 14.8% 2.49% 0.29 LLM + LORA 49.1% 14.8% 2.45% 0.30 LLM FROZEN(OURS) 50.4% 16.2% 2.71% 0.26

  42. [50]

    Data Augmentation W/OMTV 37.8% 8.7% 1.45% 0.63 W/ MTV (OURS) 50.4% 16.2% 2.71% 0.26

  43. [51]

    W. W. W

    Stage 2 EEG Encoder ENCODERFIXED44.2% 11.6% 1.75% 0.51 ENCODERTRAINABLE(OURS) 50.4% 16.2% 2.71% 0.26 C Implementation Details C.1 Model Architecture & Hyperparameters The SEMKEYframework is implemented in PyTorch, utilizing the PyTorch-Lightning library for modularity. Our arc...

    1963

  44. [52]

    Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.