acceptodds
Withdrawn from ICLR 2027

SOACT: Square-free Orthogonal Action Composition and Transport for brain-free answer prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Brain-guided language models align neural representations, but alignment alone does not specify how registered actions change a latent state or compose. SOACT predicts fixed-vocabulary answers from text using a minimal square-free quotient and registered action transport under an external action plan. It estimates nuisance-invariant moments from same-question cross-subject views using square-free products, fits polynomial action generators in a shared chart with no free intercept, composes ordered paths by substitution, and trains a text readout with the generators frozen. Independent composed-action anchors provide a proposed score-calibrated factorization test, with abstention when the evidence is unresolved. Under centered conditional noise and cross-subject independence, we show that square-free moments recover latent polynomial moments; a two-world lower bound identifies information lost by row-separable temporal objectives. We also characterize the degree identified by finite view panels and derive a scoped () versus () view-count comparison for factorized and primitive path fitting. The experiments use fixed-seed synthetic latent states, paired noisy views, action compositions, and matched comparator protocols. The reported values are evaluated under the declared experimental contracts. The task, ablation, gate, and sensitivity evaluations show the performance of SOACT and the matched non-emulating baseline under the stated protocols.

Primary area:
applications to neuroscience & cognitive science
Keywords:
BCI, neural decoding

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_vufDOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SOACT (Square-Free Orthogonal Action Composition and Transport), a framework for transferring action-conditioned latent dynamics learned from multi-subject neural observations to brain-free answer prediction. The central motivation is that conventional brain-guided language models primarily focus on representational alignment, which does not explicitly characterize how latent states evolve under registered actions or how successive actions compose. The paper also develops theoretical identifiability results, characterizes the required number of subject views, and proposes a selective factorization mechanism with acceptance, rejection, and abstention.

Critical strengths
  1. The paper explores a meaningful extension of brain-guided language modeling from static representation alignment to action-conditioned latent dynamics. Explicitly modeling how latent states evolve and compose offers an interesting perspective on transferring structural knowledge from neural observations to language models.

  2. The paper provides a relatively detailed theoretical treatment of nuisance-invariant moment estimation, polynomial action composition, identifiability, and resource requirements. In particular, the analysis of cross-subject pairing and the distinction between row-separable and full-panel estimators help clarify the information required to identify latent transitions.

Critical weaknesses
  1. First, I would like to acknowledge that this research area is not my primary field of expertise. Therefore, my assessment is mainly based on my own research experience and a critical examination of the experimental design and results. To better understand the context of this work, I also examined a closely related paper cited by the authors, “Beyond Representational Alignment with Brain-Guided Language Models for Robust Reasoning.” However, I have several substantial concerns about the present submission. In particular, the figures, including the overall architecture diagram and experimental analysis, appear to closely resemble those commonly found in conventional LLM research, without clearly demonstrating the distinctive role of brain-derived information. More importantly, although the paper positions itself within the field of brain-guided LLMs, the actual involvement of LLMs is insufficiently explained throughout the manuscript. In the experimental section, for example, the authors merely introduce a "frozen LLM" as a baseline, without adequately investigating or explaining how the proposed framework interacts with or improves the underlying language model. This disconnect between the paper's stated motivation, methodological design, and experimental validation is particularly concerning. Overall, I find it difficult to understand how the reported experiments substantiate the claimed contributions to brain-guided language modeling.

  2. The paper reports task accuracy improvements across multiple reasoning scenarios and describes evaluations involving participant-held-out records. However, it also characterizes its completed empirical validation as synthetic and acknowledges that neural subject-held-out validation and official-source comparisons remain incomplete.

  3. Although the paper is motivated by brain-guided reasoning, the authors explicitly acknowledge that the available neural artifacts do not establish the repeated, time-resolved, action-labelled observation contract required by SOACT. The Introduction also states that the reported measurements do not establish real-brain reasoning gains, and real-data validation remains outstanding.

Questions for authors
  1. Although the paper is positioned as a contribution to brain-guided language modeling, the methodology primarily focuses on latent neural dynamics, polynomial action composition, and fixed-vocabulary answer classification. In the experiments, an LLM mainly appears as a frozen feature extractor or baseline. Could the authors clarify how SOACT meaningfully advances LLM capabilities rather than simply building a classification framework on top of pretrained representations?

  2. The proposed framework does not involve free-form text generation or demonstrate improvements in the underlying LLM's reasoning process. Could the authors justify this evaluation setting and explain how the reported classification gains translate into improvements in broader language modeling or reasoning capabilities?

  3. The paper reports improvements on several reasoning classification tasks, but it is unclear whether these gains result from meaningful brain-derived constraints or from additional supervision, architectural components, and structured optimization. Could the authors provide stronger experimental evidence isolating the contribution of neural information, ideally using real brain recordings and appropriate controls?

Reviewer_qVuWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SOACT, a framework for transferring action dynamics from paired, noisy subject views to a brain-free answer classifier. Square-free cross-subject moments identify polynomial action generators under conditional independence and centering assumptions; ordered compositions then constrain a text readout while the generators remain frozen. The theoretical analysis characterizes identification and view requirements within a specified moment interface and gives a calibration-dependent deployment rule. Experiments report advantages in answer classification, operator recovery, and compositional transfer.

Critical strengths
  1. The use of paired subject views to identify shared action structure is interesting. The paper gives a fairly explicit account of what information the pairing provides.
  2. The theoretical treatment goes beyond proposing a new classifier and considers identification, composition, and the conditions needed for transfer.
  3. I appreciate the inclusion of abstention in the deployment rule. Recognizing that transfer may not be justified without calibration is a sensible aspect of the framework.
Critical weaknesses
  1. The experimental reporting is a major concern. Different parts of the paper give conflicting descriptions of whether the results come from synthetic data or real participant data, and some tables disagree on the reported values and uncertainty. I found it difficult to work out what was actually evaluated. This substantially weakens my confidence in the empirical claims.
  2. There is a substantial gap between the theoretical framework and the demonstrated system. Much of the appeal comes from the calibration-dependent guarantees, but the reported classifier experiments do not establish that the full deployment procedure works as intended. The theory and experiments therefore feel insufficiently connected.
  3. The practical case for the method is not convincing in its current form. The discussion emphasizes accuracy and savings in fitting views, while the additional calibration requirements and the consequences of abstention remain unclear. As a result, it is difficult to judge whether the proposed pipeline offers a useful advantage once all of its requirements are taken into account.
Questions for authors
  1. Can the authors clarify which experiments use synthetic data and which use real participant data, and explain the inconsistencies between the tables?
  2. Which reported results evaluate the full calibrated procedure, rather than the diagnostic classifier?
  3. What is the overall cost and prediction coverage when calibration and abstention are included?
Reviewer_qyZsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper aims to improve machine reasoning by learning from human brain activity of people solving the same questions. The proposed idea is that brain data is only used during model training, and then the model can answer new questions from text alone. SOACT works in three steps (conceptually): (1) combines brain data from different people to cancel out noise; (2) learns how each reasoning step changes the brain state; (3) maps the LLM’s latent representation to brain’s representation and trains a readout from the frozen LLM’s features to follow the same reasoning steps (or actions). The paper argues that SOACT’s design is efficient theoretically (2d vs. 2d² views) and shows that it performs better than strong comparable baselines empirically.

Critical strengths

The idea of this paper is sensible. First, combine brain data from different individuals in a way that cancels out each person’s individual noise, leaving a clean latent representation. Then, learn how each reasoning step changes the brain’s latent representation and chain multiple steps together to follow a line of reasoning. Finally, trains a lightweight mapping from LLM’s latent space to the brain's latent space to follow the same chain of reasoning, so that the model can respond to the text/previous action without needing new brain data.

Critical weaknesses
  1. The paper mentions that “experimental data comprise 20 participants, 24 task trials per participant, trial-level voxel matrices with 2,744–2,907 voxels” under section 4 Experiments. Given that the sentence mentions “voxels”, the reviewer assumes this is fMRI data. However, the paper does not explicitly state whether this fMRI data was collected by the authors themselves, or if they used data from an open dataset. If the fMRI data was collected by the authors, they should state it clearly and provide the experimental paradigm, imaging parameters, how the data was preprocessed, and why the data cannot be shared for open science practice (perhaps due to ethical considerations). If the author used data from a public database where the above necessary information is already provided, they should clearly state which dataset they used. The introduction mentions only ‘public trial-level beta artifacts available to this project’ without identifying them. One candidate dataset that could potentially be what the authors used was “Brain Correlates of Deductive Reasoning in Adults” (https://openneuro.org/datasets/ds003076/versions/1.0.1) by Booth et al. (2021), which was the dataset used by Xiao et al (2026), a key reference in this paper. The other possible dataset is the “HCP” dataset, which was mentioned twice in the whole manuscript. While the reviewer believes the “HCP” dataset refers to the Human Connectome Project dataset, the paper does not explicitly state it in the paper. The reviewer found it difficult to grasp which dataset the author used because the Booth dataset only has 17 subjects, not 20 subjects, and HCP dataset includes multiple studies, several different tasks, and thousands of subjects, and the authors did not specify which task(s) they used.

  2. In addition to not having a description for the data’s experimental design, the paper also does not have a concrete, worked example of the kind of tasks SOACT was trained to do. From Figure 1 and Section 3.4, the reviewer can infer that the task in question perhaps involves two-step or multi-step reasoning. However, the reviewer finds it difficult to understand how applicable the proposed pipeline is to understanding LLM or the human brain’s latent space when solving real world reasoning problems. The paper also does not explain how predicting “fixed-vocabulary answers from text” (stated in abstract, potentially a relevant task) can help us “specify how registered actions change a latent state or compose” (also stated in abstract).

  3. It is unclear what data allowed the authors to create Figure 2, 3, and 4. Are figures 2 and 4 based on LLM’s latent space, simulated brain data, or real brain data? Figure 3 seems to be based on real brain data from participants doing tasks, as it mentions “Permute subject names”, “Use subject means”, “Task effect” and “Neural effect”. However, as mentioned in point 1, the paper has not provided sufficient information about the dataset. Additionally, there is no code or any other supplementary material for the reviewer to verify how the result figures were created. Interestingly, the paper does seem to refer to some code and implementation. For example in the last paragraph in section 3.3, it says “run_sequential_route implements the policy.”

Reference: James R. Booth, Marisa N. Lytle, Rachna Mutreja, and Jérôme Prado (2021). Brain Correlates of Deductive Reasoning in Adults. OpenNeuro. [Dataset] doi: 10.18112/openneuro.ds003076.v1.0.1

Questions for authors

Q1. Were the results in this paper produced from real fMRI recordings, from simulation, or from a mix? Please name the dataset and explain how its experimental design is appropriate for the research question.

Q2. What exactly is an "action" for a syllogistic, transitive or relational question?

Q3. Please provide the code for how the pipeline was implemented and at least provide a concrete, reproducible example of how this paper’s approach surpasses baseline.

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

Related papers

Discussion (0)

Sign in to comment.