acceptodds
Under review as a conference paper at ICLR 2027

Useful or Merely Different? Paired-Control Evaluation for Neural Conditioning of Language Models

Abstract

Language models are typically conditioned on text alone, yet many continuous side signals can alter output distributions without the change itself being useful. We study brain-state conditioning as a control problem and introduce a paired-control evaluation principle: matched brain input should improve prediction over both removal controls (ZERO) and mismatch controls (SHUF), not just deform the next-token distribution. We quantify this distinction between generic perturbation and useful deformation with Jensen-Shannon divergence (JSD) and next-token negative log-likelihood (NLL). We evaluate this principle in two regimes. First, in controlled synthetic TVB experiments, brain conditioning strengthens with dimensionality, but only matched brain states improve NLL under the sample-wise pairing constructed in this synthetic regime. Shuffled and Gaussian-matched signals harm prediction, while freezing the brain pathway removes all effects. Second, in real MEG-MASC, we evaluate a constrained target-word likelihood task with a T5 cross-attention model over post-word MEG windows. Matched MEG improves target-word NLL over both zeroed and shuffled MEG on random, story-blocked, and subject-blocked splits, with NLL gains of roughly -. Under leave-one-subject-out evaluation, both R-Z and R-S remain below zero for multiple held-out subjects. The real effect is small and token-level, changing the top-1 prediction on about of held-out examples, and should be interpreted as controlled likelihood modulation rather than full neural decoding or improvement of a standard full-context autoregressive LM. Together, these results argue that neural-state conditioning should be judged by paired controls, revealing both a useful synthetic sanity check and a modest but real held-out MEG alignment effect.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Epistemic evaluation, Neural conditioning, Language models, MEG, Predictive reliability, Paired controls

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_F8AkOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks a simple question. When a side signal such as brain activity is fed into a language model and the output changes, is that change useful, or is the model only being perturbed? The authors propose a paired-control rule. Matched brain input (REAL) must give lower target NLL than both a removal control (ZERO) and a mismatch control (SHUF). They also separate “deformation”, measured by JSD from the ZERO distribution, from “utility”, measured by target NLL. The protocol is released as a toolkit called PairedControlBench, together with a five-item reporting standard.

Critical strengths

The MEG experiments are careful. I appreciate the range of mismatch controls, including within-subject, within-sound, local-time, circular shift, block permutation and phase randomization. The Gaussian null streams and the BERT positive control are good sanity checks on the pathway itself.

Critical weaknesses

My first concern is the novelty of the protocol itself. Removal and permutation controls are already standard in multimodal analysis. NeurIPS 2021, “Perceptual Score” measures how much a model uses a modality by permuting it across samples, which is essentially SHUF. EMNLP 2020, EMAP tests whether models really use cross-modal interactions. In neural encoding and decoding, permutation tests over mismatched segments are also routine. The paper does not discuss this line of work. I think the contribution is better framed as bringing a known two-control logic into brain-conditioned LMs and making it strict, rather than as a new principle.

My second concern is that ZERO is not a neutral control. A model trained only on nonzero brain input has never seen an all-zero input. Its behavior at zero may be arbitrary. This matters for how the results are read. At 544-D, SHUF beats ZERO, and the paper calls this a regularizer effect. A simpler reading is that ZERO is off-distribution and so a weak baseline.

Questions for authors

Were any of the models trained with brain-input dropout or zeroing, so that ZERO is in-distribution? If not, how do you rule out that the REAL vs ZERO gap partly reflects off-distribution behavior at zero? Do you have a text-only baseline with matched trainable capacity?

Reviewer_d14xOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents the MEG likelihood effect through paired-control evaluation (ZERO, SHUF). The authors consider a model is useful only if the real input outperforms both ZERO and SHUF. Furthermore, they evaluated this effect using JSD (Jensen-Shannon divergence) and target-token NLL (negative log-likelihood).

Critical strengths

[1] The authors evaluated the external model (BrainDEC) where REAL exhibits a lower NLL than SHUF but a higher NLL than ZERO. They demonstrate why the efficacy of neural inputs cannot be sufficiently judged based solely on mismatch contrasts and why ZERO must be considered alongside them.

[2] The authors utilized SHUF to design and analyze various mismatch contrasts.

[3] The authors explicitly analyzed the model's results are not coincidental through the extensive experiments with several random seeds and fold LOSO.

Critical weaknesses

[1] The authors did not sufficiently describe the specific configuration of the context-only BERT used as a positive control. Furthermore, additional explanation is required regarding the significance of this paired-control evaluation.

[2] It is difficult to clearly determine the additional predictive gain of MEG compared to the non-neural baseline based on the current results. Since the absolute NLL of MEG REAL was not explicitly reported in the evaluation and a direct performance comparison with the Gaussian models was not presented, it is difficult to determine the additional gain from the reported paired-control improvement compared to the non-neural baseline.

[3] The technical novelty of the evaluation protocol appears limited. Neural-input ablation and shuffled or noise-input controls have already been used to assess neural reliance in prior brain-to-text studies [1,2].

Reference:

  1. Ziyi Ye, Qingyao Ai, Yiqun Liu, Maarten de Rijke, Min Zhang, Christina Lioma, and Tuukka Ruotsalo. Generative language reconstruction from brain recordings. Communications Biology, 2025. (arXiv:2311.09889)
  2. Congchi Yin, Qian Yu, Zhiwei Fang, Changping Peng, and Piji Li. Rethinking cross-subject data splitting for brain-to-text decoding. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025. (arXiv:2312.10987)
Questions for authors

[1] Regarding Weakness 1, could the authors explain the specific configuration of the context-only BERT?

[2] Table 3 caution when comparing absolute NLLs between rows because each row represents a separately trained model. I would like the authors to clarify the specific reasons for this caution. If the evaluation dataset, tokenizers, target processing, and aggregation methods are identical, absolute NLLs should still allow a meaningful comparison of predictive performance across separately trained models.

[3] Could the authors provide a direct comparison between the MEG model and retrained Gaussian or constant-input baselines under identical conditions? It would help distinguish within-model paired-control improvements from additional predictive gains over non-neural baselines.

[4] [minor] - The paper contains few minor issues. For example: (a) The formatting of Table 1, (b)The given anonymous benchmark code page is not working properly, (c) Please spell out TVB at its first occurrence in the abstract for readers

Reviewer_oobXOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper evaluates neural conditioning protocols by comparing matched neural inputs against two baseline controls: completely removing the neural signal and substituting it with mismatched neural input. The authors validate this experimental setup through synthetic experiments with known ground-truth correspondence. They then present real MEG experiments demonstrating modest target-word likelihood improvements across held-out evaluations. Finally, a reanalysis of a previous fMRI checkpoint highlights the necessity of using both control baselines: while matched inputs outperform mismatched inputs, they underperform relative to removing the neural signal entirely.

Critical strengths
  • The paper’s key contribution is the empirical demonstration that relying solely on a mismatch control yields an incomplete evaluation. The fMRI reanalysis effectively shows that even when matched inputs outperform mismatched ones, conditioning can still be harmful compared to removing the signal altogether—a valuable lesson for the field.
  • The MEG evaluation includes comprehensive robustness checks, spanning multiple random seeds, cross-subject transfer, and nuisance-preserving mismatch controls. Furthermore, the authors transparently report negative results and carefully distinguish target-word likelihood improvements from overall accuracy gains, which enhances confidence in the empirical findings.
Critical weaknesses
  • The ZERO control uses the model post-adaptation to MEG data with the neural embeddings zeroed out. Consequently, it retains the adapted weight space and is not functionally equivalent to the original backbone (e.g., T5). The paper lacks comparisons against two crucial baselines:
    • The original, off-the-shelf pretrained T5 model (to establish a pre-adaptation benchmark).
    • A matched model trained on the exact same target-word prediction task without any neural inputs.

These baselines address distinct questions: the former indicates performance prior to fine-tuning, while the latter isolates the specific value added by neural conditioning over text-only adaptation. Without them, the current empirical results demonstrate an advantage over input ablation and mismatch within an adapted model, but fail to prove whether neural conditioning offers a broader benefit over standard text-only modeling.

Questions for authors
  • Were zeroed (or ablated) neural inputs presented to the model during training (e.g., via modality dropout or input masking)? Clarifying this would help evaluate whether the baseline represents a familiar or an out-of-distribution input condition for the adapted model.
  • How does the MEG-conditioned model perform compared to the original pretrained T5 baseline and an otherwise identical model trained on the target-word objective without neural inputs? Evaluating all models under the same evaluation protocol would help disentangle the benefits of task-specific adaptation from the true signal contribution of neural conditioning.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.