acceptodds
Under review as a conference paper at ICLR 2027

NMCNet: A Cortex-Guided EEG-EMG Fusion Framework for Cross-Subject Decoding of Mandarin Initials and Finals

Abstract

Brain-computer interface (BCI) speech decoding offers a promising communication pathway for individuals with severe speech impairments. However, Mandarin initial and final decoding remains underexplored, and existing EEG-EMG fusion methods rarely encode an asymmetric interaction motivated by cortical control and articulatory muscle responses. We propose NMCNet, a cross-subject EEG-EMG fusion framework for audible and silent speech decoding. NMCNet integrates Cortical Drive Gating, which uses EEG features to modulate EMG representations, with Feedback Cross-Modal Fusion, which reintegrates the resulting muscle-response information into the EEG representation. A complementary multi-frequency branch captures frequency-specific dynamics. To improve cross-subject generalization, we further introduce a cross-subject-enhanced supervised contrastive loss that emphasizes same-class samples from different subjects. A Fisher-based channel-selection strategy reduces the EEG input to 15 channels while outperforming the full channel configuration, supporting lightweight acquisition. Across ten participants, NMCNet achieves the best performance under silent speech, with a mean accuracy of 70.25% across the initial and final tasks, each comprising three broad articulatory categories. In a fine-grained setting that treats each individual initial or final within its broad category as a separate class, NMCNet achieves macro-averaged F1 and recall scores of 73.48% and 74.35%, respectively. Ablation studies support the effectiveness of the proposed architectural components and cross-subject learning strategy. These results demonstrate the potential of lightweight cross-subject Mandarin decoding and silent-speech BCIs for individuals with anarthria, severe dysarthria, and other severe motor speech impairments.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
EEG, EMG, multimodal learning, cross-subject generalization, silent speech

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_njCWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NMCNet, a multimodal framework for cross-subject classification of Mandarin initials and finals using synchronized EEG and EMG. Its central architectural component is an asymmetric interaction mechanism: Cortical Drive Gating uses EEG features to modulate EMG representations, and Feedback Cross-Modal Fusion subsequently uses the gated EMG representation to update the EEG representation. The model additionally incorporates a multi-frequency branch and a supervised contrastive objective that upweights same-class pairs from different subjects. The authors evaluate the method on a self-collected dataset of ten healthy participants performing audible and silent articulation. Experiments include three-class broad-category decoding and six within-category fine-grained tasks. The reported results are promising, especially for final classification. However, important uncertainties concerning supervised channel selection, model selection, phonetic labels, and baseline comparability currently limit confidence in the main conclusions.

Critical strengths

S1. Relevant problem and motivated architecture. The paper investigates a meaningful multimodal decoding problem and explicitly considers the different roles of EEG and EMG rather than treating them as interchangeable inputs. The asymmetric fusion mechanism provides a coherent architectural hypothesis for combining these modalities. Evaluating this hypothesis in a cross-subject Mandarin articulation setting is a worthwhile research direction. S2. The ablation studies on the fusion architecture are quite comprehensive. Beyond removing individual modules, the authors also compare reverse fusion and symmetric fusion (with parameter matching). These experiments are more insightful than simple module removal for determining whether the direction of information interaction is significant, and whether performance gains stem merely from an increase in the number of parameters. Furthermore, the single-modality control experiments provide evidence for the complementary nature of EEG and EMG data. S3. Performance improvements are notably significant for some tasks, and results are provided on a per-subject basis. As shown in the main results table, the model demonstrates a much greater advantage in classifying final rhymes than in classifying initial consonants. The per-subject results in the appendix also help determine whether the average performance has been disproportionately influenced by individual subjects. The authors’ inclusion of data reflecting variations across different tasks and subjects is a positive aspect of this experimental report.

Critical weaknesses

W1. The task definitions and the scope of the decoding claims need greater precision The fine-grained experiments appear to classify individual initials or finals within a preselected phonetic group, rather than jointly classifying all target initials or finals without knowing the group. These are different evaluation settings. Consequently, the average macro-F1 across the within-group tasks should not be interpreted as the overall performance of a complete initial/final recognition system. The manuscript should clearly state whether ground-truth group membership is assumed at evaluation time and whether a practical system would require an additional group classifier. A related concern is that the initial and final labels are derived from the same finite set of syllable trials, while the model processes a signal segment covering the full articulation. LOSO evaluation tests generalization to unseen subjects, but it does not by itself distinguish between learning compositional initial/final representations and learning whole-syllable patterns associated with the target labels. The authors should provide the complete mapping between stimuli, initials, finals, and coarse categories. A held-out syllable-combination evaluation would strengthen the claim that the learned representations generalize beyond the observed combinations. If compositional generalization is outside the paper’s intended scope, the claims should instead be limited to cross-subject classification within the evaluated stimulus inventory. W2. The phonetic grouping needs clarification In Mandarin Pinyin, j is an affricate, not a sonorant. Why is it placed in the sonorant group rather than the fricative/affricate group? Please provide clear phonetic criteria for the grouping and explain this assignment. Please also clarify whether this is a mistake in the manuscript or whether the same grouping was used in the experiments, since the latter would affect the training labels and the interpretation of the results. W3. The independent contribution of asymmetric fusion is not fully separated from the training recipe NMCNet combines asymmetric fusion, multi-frequency processing, normalization, Mixup, noise augmentation, a classification objective, and supervised contrastive learning. Although the reversed and parameter-matched symmetric variants are useful, the manuscript should more clearly document whether the external baselines use comparable inputs, preprocessing, augmentation, loss functions, and hyperparameter-search budgets. A particularly informative comparison would preserve the encoders, multi-frequency branch, losses, and training procedure while replacing CDG and FCMF with direct concatenation or a conventional gating mechanism. This would provide a cleaner estimate of the benefit attributable specifically to the proposed asymmetric interaction.

Questions for authors

Q1. Can the authors provide the complete stimulus-to-label mapping and clarify the intended scope of the fine-grained results? Are unified or cascaded results without ground-truth group membership available? Q2. What does j denote in the phonetic grouping, and does the grouping shown in Figure 2 correspond to the labels actually used for training and evaluation? Q3. Were channel ranking, channel-count selection, hyperparameter tuning, and checkpoint selection performed exclusively using training subjects in every outer LOSO fold? Which reported results correspond to fold-wise training-only channel selection?

1 comment

We thank the reviewer for the most detailed assessment of this submission; the critique shaped the revision, and every experiment has been re-run under a strictly leak-free protocol.

W3. We ran the requested control. Holding the encoders, the multi-frequency branch, Mixup, Gaussian noise, the contrastive loss and the optimiser settings fixed, replacing the entire CDG+FCMF block with a plain channel-wise concatenation reduces κ from 0.396 to 0.294, and the resulting model falls below DeepConvNet (0.329); the gain is not attributable to the encoder alone, and under the same training recipe this supports the contribution of the fusion block. We also re-examined the cross-subject reweighting: κ varies by less than 0.03 across γcs = 1.0–2.0 with no consistent direction (paired γcs = 1.0 vs 1.75: +1.69 points on initials, p = 0.24; −1.92 on finals, p = 0.39). We withdraw it as a contribution and report it as a sensitivity check.

Q3. All selection now happens within the training subjects. The channel ranking for outer fold k is computed from the nine source subjects only, using four condition-specific ANOVA p-values combined by Fisher's method; k = 15 is fixed a priori; the stopping epoch is chosen by an inner five-fold split of those nine subjects, after which the model is retrained on all nine and tested once. The held-out subject contributes to no stage. Ranking stability is Jaccard 0.62 ± 0.10 (4.2× chance). Fifteen channels and all 59 give similar observed results (65.27% / κ 0.396 vs 64.47% / κ 0.393); no formal equivalence test was performed, so we describe this as similar rather than equivalent. Channel selection is reproducible but not performance-critical. The revision will include the full per-fold ranking.

W2 / Q2. The reviewer is right that j ([tɕ]) is an affricate and not a sonorant. We merged fricatives and affricates into one class, giving Plosive (b, d, g, p, t), Fricative/Affricate (zh, ch, c, sh, h, f, s, j) and Sonorant (m, n, l, y, w, zero-onset). All results are recomputed; the revised figure will show the labels used for training and evaluation.

W1 / Q1. The fine-grained results are oracle-group analyses given the true coarse category, not system-level performance, and the table will be relabelled accordingly. The complete stimulus-to-label mapping is given below; each syllable has 16 trials per participant. Category rules: Plosive = b, d, g, p, t; Fricative/Affricate = zh, ch, c, sh, h, f, s, j; Sonorant = m, n, l, y, w, zero; Monophthong = a, i, u, e, o; Diphthong = ai, ia, ui, ao, ou; Nasal coda = an, ang, eng, uan.

SyllableOnsetFinal
mama
baba
jiji
yaya
yiyi
fanfan
gege
cici
sheshe
jiajia
mimi
juju
huihui
fufu
aizeroai
dada
pipi
momo
wangwang
youyou
nini
fengfeng
yuyu
changchang
tuitui
huanhuan
lulu
suisui
maomao
zhuzhu

We do not claim generalisation to unseen onset–rhyme combinations. What the present results establish is cross-subject classification on a fixed 30-syllable inventory (two articulation modes × two phonological levels over six groups, 16 trials per syllable per participant); we have not tested whether the model generalises to a syllable whose specific onset–rhyme combination was absent from training.

Such a test is feasible with this corpus and is the natural next step: holding out a syllable such as ma leaves mi, mo and mao in training (so the onset m is still observed) and ba, da, ya (so the final a is still observed), which yields an unseen-combination test at the level of both coarse categories. We have not yet run it; a leave-one-syllable-out variant of the LOSO protocol is the immediate next step, and extending this to a combinatorial stimulus set and to open-set decoding is future work.

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

This work mainly focuses on improving the performance of brain-computer interface (BCI)-based speech decoding, especially the decoding of Mandarin initials and finals. To this end, the authors propose a cross-subject EEG–EMG fusion framework called NMCNet. Specifically, NMCNet introduces a temporal branch and a multi-frequency branch to extract diverse representations. NMCNet also introduces Fisher-based Channel Selection to improve efficiency and Cross-Subject-Enhanced Supervised Contrastive Loss to enhance generalizability. Extensive experimental results demonstrate the effectiveness of the proposed method.

Critical strengths

The research problem is clearly defined and remains to be addressed. It has significant value for real-world applications.

NMCNet achieves strong results on the speech decoding task, especially in cross-subject settings.

The authors conducted extensive experiments to validate the effectiveness of the method. The paper is well organized and easy to follow.

Critical weaknesses

The motivation and definition of asymmetric fusion are insufficiently explained.

The technical novelty is limited, as the method largely combines existing techniques without clearly articulating its novel contributions or the motivation for the proposed combination.

The effectiveness of certain design principles included in the NMCNet framework remains uncertain.

Questions for authors
  1. The authors need to explain why asymmetric fusion is necessary. They are also advised to clarify exactly what asymmetric fusion is and how it differs from conventional fusion strategies.

  2. Limited Novelty: While this work demonstrates technical soundness, its novelty appears limited. The proposed method mainly combines existing techniques and lacks sufficient motivation.

  3. In Table 2, although NMCNet generally performs well on the speech decoding task, its performance improvement on the Silent-Initial Decoding task is very limited (only +0.45 percentage points). This raises concerns about whether the proposed fusion framework can effectively capture rapid articulatory transitions.

1 comment

We thank the reviewer for the focused critique. We agree that the original manuscript did not explain why the asymmetric ordering is necessary, and that the reported margins did not support some of its claims.

Q1. The motivation is functional: EEG carries the cortical drive that gates articulation, EMG the executed muscular activity. The design uses the EEG representation to modulate EMG features (CDG) and returns the gated EMG to update the EEG representation (FCMF), so the two modalities interact in a specific order rather than being merged symmetrically. To test the contribution of the fusion block under an identical recipe, we replaced the whole fusion block with plain concatenation: κ falls from 0.396 to 0.294 and drops below DeepConvNet (0.329).

Q2. The contribution is conceptual before it is architectural. We accept that the ingredients are established techniques and have removed the "novel combination of modules" framing. The revised contributions are: (i) a physiologically motivated fusion ordering, supported by the concatenation control above; (ii) a task-level contribution — cross-subject multi-class decoding of Mandarin initials and finals as phonological units, previously studied only within-subject and in binary form, under both audible and silent articulation over a fixed 30-syllable inventory, with the first leak-free benchmark; and (iii) a characterisation of where the gain arises, namely on sustained finals (Δaccuracy = +20.48 and +15.21 points) rather than on transient initials (+1.13 and +2.73, not significant).

Q3. The +0.45-point margin was computed under the previous protocol, which selected the stopping checkpoint on the held-out subject. Each participant performs both tasks, so the two conditions are two measurements of the same participant, not independent experiments; treating them as replications would count each participant twice. Our inferential unit is therefore the participant, consistent with the subject-wise averaging used throughout. At that level NMCNet is significantly better than DeepConvNet: Δaccuracy = +9.89 points over four conditions (10/10 participants, p = 0.0020) and +8.97 over the two silent conditions (10/10, p = 0.0020). We do not overstate the decomposition: on silent-initial alone the margin is +2.73 points and is not significant (7/10, p = 0.090); the advantage is carried by the final conditions. We no longer claim a silent-initial advantage, and the revision will report per-condition significance so that this decomposition is visible.

Reviewer_wRz6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce NMCNet, which is an EMG-EEG fusion model for cross-subject classification of Mandarin initials and finals. They explore this on a dataset of 10 speakers during audible and silent articulation. The main components of the architecture are separate encoders for each modality, an asymmetric gating mechanism, and the use of a subject-aware contrastive objective as an auxiliary loss. Results are promising (in particular, on finals), but I have concerns about the evaluation protocol and whether the evidence at this stage supports the proposed cortical interpretation.

Critical strengths
  • The application is relevant, and I liked the approach and evaluation on unseen participants, addressing the cross-subject generalization problem, which is still open in the field.
  • Ablations are sound, including modality ablation, multimodal baselines, and checks of the symmetric vs. asymmetric fusion mechanism.
  • The model generalizes well across participants (mostly on finals) and also separates well across silent conditions.
Critical weaknesses
  • One of the main things that is unclear to me is the validation protocol. Channel selection uses class labels, but does this exclude the held-out participant? Is there a defined validation/test split for hyperparameter selection and early stopping? If so, it would be helpful to clearly report it in the manuscript; otherwise, the results might be optimistic.
  • The cortical interpretation is not sufficiently proven here. Participants are looking at letters and then voicing them. Even after ICA cleaning, it is difficult to rule out that the EEG contribution is not driven by residual visual perception (reading) or muscular contamination. The fact that the network outperforms the baselines speaks to the quality of the architecture and not to the cortical vs. other types of contributions to the EEG data. I would tone down this part and make the limitations clearer.
  • The methodological novelty is limited and incremental. The method works, but it combines established components.
  • It would have been nice to see some examples of generalization across unseen sounds (initials or finals), if possible.
Questions for authors

Please see the weaknesses above. On top of that, I’d like to ask the authors:

  • Were channel ranking, channel-count selection, hyperparameter tuning, and early stopping performed without using the held-out participant’s labels? Please report the fold-wise channel-selection results and describe the validation split.
  • What evidence distinguishes cortical information from residual muscle activity and visual/lexical processing in the EEG contribution? Would the gains persist under stricter artifact controls or evaluation on unseen syllables?
3 comments

We thank the reviewer for the positive assessment of the application relevance and of the ablation design, and for the methodological scrutiny that led us to re-run the entire evaluation under a strictly leak-free protocol.

1. Validation split and channel selection. Both were unspecified and are now fully defined. The channel ranking for outer fold k uses only the nine source subjects (four condition-specific ANOVA p-values, Fisher-combined); the held-out subject contributes to no stage. The inner loop is a five-fold split of those nine subjects, used only to choose the stopping epoch, after which the model is retrained on all nine and tested once. k = 15 and all hyperparameters are fixed a priori from prior work. Ranking stability is Jaccard 0.62 ± 0.10 (4.2× chance), and fifteen channels show similar observed performance to all 59 (65.27% / κ 0.396 vs 64.47% / κ 0.396).

2. Cortical interpretation. We agree that the previous wording overstated the conclusion. The paradigm is a cue–articulation sequence: participants read a character and then articulate it, so the cue-evoked response is part of the signal available in the intended use case, since the target population retains reading ability. The revised text will describe the decoded information as the response to a cue–articulation sequence rather than an isolated motor command, will state that the system does not apply to patients with impaired reading, and will confine the architecture-level claim to what the concatenation control supports.

3. Novelty. The contribution is conceptual before it is architectural: the model encodes a specific physiological asymmetry — cortical drive gating the articulatory muscles — rather than merging modalities symmetrically. The revised contributions are (i) this physiologically motivated fusion ordering, supported by the concatenation control (κ 0.396 → 0.294, below the DeepConvNet baseline of 0.329); (ii) a task-level contribution, cross-subject multi-class decoding of Mandarin initials and finals under audible and silent articulation over a fixed 30-syllable inventory, with the first leak-free benchmark; and (iii) a characterisation of where the gain arises — on sustained finals (Δaccuracy = +20.48 and +15.21 points) but not on transient initials (+1.13 and +2.73, not significant).

4. Generalisation to unseen syllables. The study is closed-set over a fixed 30-syllable inventory. What the present results establish is cross-subject classification on that fixed set; we have not tested generalisation to a syllable whose specific onset–rhyme combination was absent from training. Such a test is feasible here — holding out ma leaves mi/mo and ba/da in training — but we have not run it; a leave-one-syllable-out variant of the LOSO protocol is the immediate next step. Extending this to a combinatorial stimulus set and to open-set decoding is future work.

Thank you for the detailed response and for rerunning the evaluation under the clarified protocol. The source-only channel selection and inner validation procedure substantially address my concerns about potential leakage. I also appreciate the more explicit limits on the physiological interpretation and on generalization beyond the fixed syllable inventory.

However, my remaining concerns are not fully resolved. Describing the task as a cue–articulation sequence acknowledges the contribution of cue processing, but does not establish whether the EEG gains reflect cortical information rather than residual muscular contamination. Likewise, the concatenation control supports the utility of the proposed fusion mechanism, but does not establish that it captures the claimed physiological directionality from cortical activity to muscle activity. This distinction remains relevant because that interpretation is still central to the argument for conceptual novelty.

Overall, the response improves the methodological clarity and better defines the scope of the findings. Nevertheless, my concerns about the strength of the physiological evidence and the limited methodological novelty remain. I therefore maintain my rating of 2 (weak rejection), even though, in principle, if other reviewers raise their scores, I’m not strongly against the acceptance of this work.

We thank the reviewer for the positive assessment and for raising this point. We have analysed the concern directly, and the result is reported below.

To test whether the decoding depends on sensory-feedback sites, we removed all visual (O1, O2, Oz, PO3–PO8, POz), auditory/temporal (T7, T8, TP7, TP8, FT7, FT8, P7, P8) and frontal-polar (Fp1, Fp2, Fpz, AF3, AF4, AF7, AF8) electrodes from the candidate pool. The fold-wise ranking then selected its fifteen channels from the remaining candidates, so the input dimension, the number of channels and every hyperparameter were unchanged.

Accuracy (%)

ConditionAll channelsRemovedΔp
Initial66.1065.08−1.020.414
Final67.7167.31−0.400.922
Mean66.9166.20−0.710.625

Cohen's κ

ConditionAll channelsRemovedΔp
Initial0.4570.434−0.0230.232
Final0.3840.402+0.0180.846
Mean0.4210.418−0.0030.922

Paired tests over the ten participants (two-sided Wilcoxon signed-rank). No comparison approaches significance. No occipital or temporal electrode is selected in any fold; the selected set remains central and parietal.

On the physiological framing: the ordering of the two modalities is motivated by an established account of speech motor control rather than by our own experiments. Human speech sensorimotor cortex carries a somatotopic representation of the articulators (Bouchard et al., Nature 2013) and encodes articulatory kinematic trajectories during production (Chartier et al., Neuron 2018); the DIVA model formalises the transformation from cortical speech-sound targets to articulator commands (Guenther et al., Psychological Review 1998; Tourville & Guenther, Language and Cognitive Processes 2011); and corticomuscular coherence shows that motor cortex and the innervated muscle are coupled, both in the limb system (Conway et al., Journal of Physiology 1995) and in the orofacial system, where coherence peaks at 2–3 Hz, the spontaneous rhythmicity of speech (Ruspantini et al., Journal of Neuroscience 2012).

We cite this literature as the prior our design builds on, not as evidence that our model implements the mechanism. The ablation above likewise constrains where the information comes from without establishing a cortical origin. Accordingly we describe the framework as cortex-inspired rather than cortex-guided, and we do not claim causal directionality from cortex to muscle. These results, together with the per-participant values, will be presented in the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank all reviewers. Since submission we have re-run every experiment under a strictly leak-free protocol: channel ranking is computed per fold from the nine source subjects only, the number of channels is fixed a priori (k = 15), hyperparameters are fixed from prior work, and the stopping epoch is selected by inner five-fold cross-validation over source subjects. The held-out subject is used for evaluation only.

Contribution framing. The design is not a generic multimodal fusion. It follows a specific physiological asymmetry: articulation is initiated by cortical commands that drive the articulatory muscles, so EEG carries the cortical drive and EMG the executed muscle activity. NMCNet encodes that ordering — the EEG representation modulates the EMG representation, and the gated EMG is returned to update the EEG representation — instead of merging the two modalities symmetrically. The task is also new in this setting: cross-subject decoding of Mandarin initials and finals as phonological units (rather than lexical tone), under both audible and silent articulation, over a fixed 30-syllable inventory. The concatenation control shows that removing this ordering removes the advantage: κ falls from 0.396 to 0.294, below the DeepConvNet baseline.

Leak-free re-evaluation. Accuracy (%) and Cohen's κ, mean over ten held-out subjects.

MethodA.I.A.F.S.I.S.F.Avg accAvg κ
NMCNet61.4665.8166.1067.7165.270.396
DeepConvNet60.3345.3363.3852.5055.390.329
DRDA53.5856.8953.0957.4655.260.232
EEGNet47.8145.0847.9442.2145.760.158

Subject-wise accuracy under silent articulation (mean of the two silent conditions, %), replacing the values previously reported in Table 3:

MethodS1S2S3S4S5S6S7S8S9S10Avg
NMCNet46.7761.2569.6965.3171.2569.0671.8871.1575.4267.2966.91
DeepConvNet39.3848.8662.8153.1265.3153.9562.1962.1966.7764.7957.94
DRDA42.5847.6157.4556.5956.4157.8758.3656.8360.6258.4455.27
EEGNet—42.0844.5945.5244.3839.9049.1748.0246.7748.0245.38

NMCNet is highest in all ten participants against every baseline, so the advantage is not driven by a few subjects. One EEGNet fold is unavailable for silent-initial.

Channel count. Under the leak-free protocol, the 15 fold-wise selected channels and all 59 channels show similar observed performance:

ChannelsA.I.A.F.S.I.S.F.Avg accAvg κ
15 (fold-wise)61.4665.8166.1067.7165.270.396
59 (all)62.1565.3863.9666.3864.470.396

The difference is −0.80 points of accuracy and −0.003 of κ. The ranking itself is reproducible: pairwise Jaccard 0.62 ± 0.10 across folds, 4.2× the chance level. Channel selection is therefore a reproducible but not performance-critical preprocessing step, and k = 15 was fixed a priori rather than selected on the test set.

Fusion ablation. We replaced the entire CDG+FCMF block with a plain channel-wise concatenation of the two modality features, holding the encoders, the multi-frequency branch, Mixup, Gaussian noise, the contrastive loss and the optimiser settings fixed. This reduces κ from 0.396 to 0.294 (Δκ = −0.102). The concatenation variant also falls below DeepConvNet (κ = 0.329), so the improvement is not attributable to the encoder alone; under the same training recipe, this supports the contribution of the fusion block.

Withdrawal. In the revision we fix γcs = 1.75 uniformly across all four conditions, replacing the condition-specific values used previously (1.35 / 1.75 / 1.75 / 1.40). Under this fixed setting the previously reported effect of cross-subject reweighting does not replicate: κ varies by less than 0.03 across γcs = 1.0 … 2.0 with no consistent direction, and the paired comparison of γcs = 1.0 versus 1.75 gives +1.69 points on initials (p = 0.24) and −1.92 points on finals (p = 0.39). We therefore withdraw cross-subject reweighting as a contribution and report the analysis as a hyperparameter sensitivity check.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.