acceptodds
Under review as a conference paper at ICLR 2027

INCEPT: Infusing Novel Concepts for Explaining Pretrained Tabular models

Abstract

The Platonic Representation Hypothesis suggests that sufficiently capable models converge to similar internal representations regardless of architecture or training. Using embeddings from the TabArena benchmark to assess geometric convergence of tabular foundation models, we find transformer-based tabular foundation models (TFMs) cluster tightly, with hypernetworks and LLM-based models sitting apart. To explain the predictive gap between two TFMs, we train a sparse autoencoder (SAE) per model and identify their unmatched concepts. Two interventions are tested: ablation zeros unmatched concepts out of the stronger model's embedding; transfer maps them into the weaker model's embedding. Across 15 model pairs and 51 TabArena datasets, ablation closes 93% of the stronger model's per-row advantage, and transfer recovers 90% of the gap. The transferred concepts lie roughly half off the weaker model's existing manifold, revealing structure the recipient does not use but can absorb.

Primary area:
interpretability and explainable AI
Keywords:
sparse autoencoders, mechanistic interpretability, tabular foundation models, representation convergence, causal interventions

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_W7YsOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The submission investigates the "Platonic representation hypothesis" in the context of tabular foundation models: To what extent do different tabular foundation models with different architectures and different training data converge to similar internal representations. It trains sparse autoencoders (SAEs) on each model, matches SAE features across models and evaluates whether ablating matched and unmatched features explains differences in model performance.

Critical strengths
  • I find the general research question interesting: To what extent does the Platonic representation hypothesis apply to tabular foundation models given their diversity of architectures and training data
Critical weaknesses
  • The presentation of this submission is poor: It uses dense, jargon-heavy language that often omits necessary context, making the reader guess what the possible intended meaning of a sentence could be. This is particularly problematic in terms of understanding the experimental setup and results. There are many statements throughout the manuscript of which I cannot make sense unfortunately (see questions for examples).
  • It is difficult to assess the technical merits of this submission with certainty given the severe presentation issues but I am not sure why it would be surprising that "infusing missing concepts" reduces the gap in performance between two models (see questions).
Questions for authors
  • Why would it be surprising that "infusing missing concepts" reduces the gap in performance between two models? What else could explain the gap in performance?

Transfer requires a map between the two models’ embedding spaces. We estimate a global linear map M from matched concept pairs via ridge regression, filtering landmark correlations above the random-SAE noise floor (details in Appendix F.8).

  • I am unfortunately not able to infer what was done here. Could you please explain in simple words how you use the matched concept pairs from two models' SAEs to establish a map between embedding spaces? On what is the ridge regression run? Why can't you use the SAEs directly for this?

We apply Leiden community detection Traag et al. (2019) to partition this graph into tightly-connected communities, each representing a single shared concept.

  • This statement seesm to apply that this procedure guarantees a single shared concept per partition? Why is this the case?

  • I am missing a comparison of how the pretraining data affects representational similarity across models (in addition to how architecture affects it). The introduction highlights the diverse data tabular models are trained on and I was expecting such an analysis as a consequence.

To find the layer where the optimal representations could be found for each model, we set output hooks at every intermediate layer to cache intermediate activations for every dataset.

  • Optimal for what? The paragraph on "Embedding extraction" was hard to parse, could you re-explain to me based on what criterion you select the intermediate layer used for the ensuing model comparison?

The CKA analysis establishes that models share geometric structure, but does not reveal what that structure encodes or why models differ in performance.

  • I am confused by this statement, which analysis results does this refer to at that point in the paper?

Matryoshka mitigates concept splitting, archetypal constraints improve cross-seed stability, and top-k sparsity reinforces both.

  • This sentence seems odd, are you stating that top-k sparsity reinforces concept splitting?

Since models have different embedding dimensions (192 to 768), their decoder weight vectors are not directly comparable.

  • CKA works in this setting, it is not clear to me why it cannot be used to calculate "Activation-based similarity" (the title of that paragraph)

The number of accepted concepts K is not preset: the per-row greedy admits concepts in importance order while each moves the recipient toward S, and stops when gap closed reaches a tolerance or no remaining concept improves it.

  • The submission suffers from poor writing, this is an example sentence that illustrates this. It takes several readings to guess an intended meaning.

Tabular foundation models present a more challenging test: architectures span transformers, hypernetworks, and graph neural networks; training data ranges from synthetic priors to real tables; and there is no shared “Common Crawl” equivalent. Building on geometric tools—CKA and Procrustes alignment—we go beyond that comparison with concept-level causal analysis, examining what models learn and why they differ in performance.

  • What is "that comparision"? Is there a sentence missing?

  • What is a TPE sweep (L198)? This acronym has not been introduced.

  • The symbols in equation (2) are not defined. What is fi(A)f_i^(A)?

  • How/where is "gap closed gc" used in Algorithm 1?

  • I am unable to decipher Figure 1a. Can you explain in simple words what is inteded to show?

The interventions are a causal diagnostic, not a deployment method: they take per-row knowledge of which model is stronger as an experimental control

  • I don't understand what that means

Minor remarks

  • There are several instances where citations are missing their parentheses (citet instead of citep), e.g. L167
Reviewer_jvEuOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses a problem of: Can the differences in prediction performance among different Tabular Foundation Models be attributed to the different semantic concepts contained in their internal representations? The author compared the representation geometries of 8 TFMS on 51 TabArena datasets and conducted concept-level interventions on 6 of the models.

Critical strengths

The paper addresses an interesting and relatively underexplored problem of understanding representation differences among tabular foundation models. Combining SAE-based concept matching with ablation and transfer interventions provides a promising framework for analyzing cross-model differences. The evaluation across multiple models and 51 datasets, together with random-SAE controls, offers encouraging empirical evidence.

Critical weaknesses

The same sample was used for concept selection and evaluation. The intervention direction of each test row is selected based on the predicted changes of that row, so it is problematic to distinguish the truly stable explanatory features from the post hoc fitting for the current prediction results.

The proposed concept transfer relies on a global ridge regression mapping between the representation spaces of different models. However, the authors acknowledge that this mapping achieves a negative cross-validated, indicating poor alignment quality. This raises concerns about whether the transferred directions genuinely preserve the semantic information of the original concepts or merely serve as effective perturbations selected by the oracle-guided greedy procedure.

Questions for authors

How can the authors ensure that the reported gap closure reflects genuine explanatory concepts rather than post hoc optimization on individual test samples?

Given the negative cross-validated of the ridge regression mapping, what evidence supports the claim that the transferred directions preserve meaningful semantic information?

Are activation-level correspondence and semantic-level correpondence the same concept? It is not clear they share consistency and the evidence is not sufficient.

Reviewer_X8p3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates Platonic Representation Hypothesis in tabular foundation models and proposes two investigation mechanisms which it weirdly calls ablation and transfer. Particularly, it investigates why pretrained tabular models make different predictions. It uses SAEs to extract concepts, and identifies concepts present in one model but missing from another. The paper reports that 2-5 'unmatched' selected SAE concepts in a strong model is the main driver of the performance gap. They also find that transformer-based tabular models have more similar representations than models from other architectural families.

Critical strengths
  • Mehcanistically studying tabular models is an important direction and understanding discrepancies can push the field forward.
  • The paper seems to examine 51 datasets and 15 model pairs, that is comprehensive.
  • Identifying a small number of selected concepts to explain the gap presents an interesting finding.
Critical weaknesses

Abstract/Intro: First, do authors understand what 'ablation zeros unmatched concepts out of the stronger model's embedding' would mean without context? if so please tell me. Second, similarly 'transfer maps them into the weaker model's embedding' => this is true for any embedding. take any weak and strong model, I can provide an arbitrary map that maps one to the other. So it says nothing. With two meaningless sentences, the abstract fails to say anything. 10 points for Claude.

Similar things happen throughout the paper at times, so the authors could do a little better job to filter than the current state. For instance, the introduction says "model’s gap is recoverable" : basically anything is recoverable if you do something to it, or "Transfer also never degrades the recipient": totally unclear what this sentenc means. Pease avoid such adjectives. Essentially you are simply wasting space with these, instead of giving us good information and insightful discussion. In general I find the paper hard to read. As such, even if the paper is strong, it would get negative reviews.

Method/Experiments:

  • Why are we interested in closing the gap between TabFMs in the first place? I find that the paper lacks a clear motivation.
  • I think this paper should also consider the 'Aristotalean' version of the hypothesis, where local structures are meant to align. A similar idea led to Schnaus et al. to come up with a vision-language unpaired alignment: https://arxiv.org/pdf/2610.09411.
  • In general, it is good to have .
  • I'm not sure if the interventions are to be useful or they are for the sake of interpretability. Using some form of higher-order concepts is a classical, good idea, but it's not new in itself. So, it can lead to some better unpaired alignment perhaps? And I expect that this could have been investigated.
  • Using SAEs to represent concepts present additional challenges and it gets harder to detach wether the gains are true or stem from SAE imperfections.
Questions for authors

see weaknesses.

Comments on the paper

2 comments

Thank you for your response to my questions. I appreciate the additional hypotheses you added to attempt to explain what might might explain the closing of the gap to the stronger model.

As I highlighted in my initial review, a central concern for this submission is its poor presentation, specifically the dense, jargon-heavy language used throughout, that often omits the context that would be necessary to understand what is actually being done. In the same vein, it took me considerable effort to parse the responses to my questions. As it stands, I cannot confidently recommend this paper for acceptance and will maintain my score. That being said, I think the research direction is interesting and I would strongly encourage the authors to carefully rewrite the manuscript to make it (human) readable for re-submission at a future conference.

Q2 (ridge map; why not SAEs directly). An unmatched concept has no counterpart in the weaker model's dictionary, so it cannot be expressed through the weaker model's SAE. Each matched pair supplies one paired decoder direction, and we fit a ridge map on those pairs (§3.3; Appendix F.8). The ridge map then lets us infer where the unmatched concept will sit when the concept is transferred to the weaker model's embedding space.

This made me wonder if there is some way to characterize this "weight edit" compared to fine-tuning the model for the particular error that the weight edit fixes. Are the respective updates aligned to each other? Is this effectively like training the model?

Thank you for the reply and for the question. Your question points to the logical next step - ok, so what do we do with this information? Our hope with the submission was that it would resonate with the TFM community - and as they are the experts, they would be best placed to act on our current diagnostics.

Our corpus of 46,000 transferred rows might be a start for anyone wanting to embark on "weight editing" or fine-tuning.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.