acceptodds
Under review as a conference paper at ICLR 2027

Beyond Linear Representation Hypothesis: Non-Linear Activation Steering in Text-to-Image Models

Abstract

Mechanistic interpretability often relies on the Linear Representation Hypothesis (LRH), which assumes that high-level concepts are encoded as linear directions in activation space. Yet a natural visual concept does not necessarily require a linear visual transition: between sunny and stormy lies an intermediate weather state such as a sky with a few white clouds, not simply a weaker storm; between a caterpillar and a butterfly, the progression is not a caterpillar with continuously growing wings. This raises the question of whether such true intermediate states are also represented nonlinearly by the model. Indeed, when we prompt text-to-image models directly for intermediate attributes, their activations rarely fall along the straight direction connecting the endpoints. Therefore, we propose KANSteer, which models concept traversal as a curve passing through its intermediate states. Seeking a representation that is both simple and interpretable, we propose to use Kolmogorov-Arnold Networks (KANs), which provide a one-dimensional coordinate whose learned functions define the trajectory. This allows the steering direction to vary along the concept while preserving an interpretable representation. Across several concepts and text-to-image diffusion transformers, we find that their activation trajectories substantially deviate from straight lines, and that KANSteer provide a closer fit and smoother traversal of intermediate attributes than linear steering.

Primary area:
generative models
Keywords:
mechanistic interpretability, diffusion models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_B3eeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes KANSteer, a nonlinear activation steering method for text-to-image diffusion models. By modeling concept transitions as one-dimensional curved trajectories rather than fixed linear directions, the method enables smoother traversal of intermediate visual attributes. The experimental results demonstrate consistent improvements over linear steering across several concepts and model configurations.

Critical strengths

The paper is clearly motivated and well presented. The idea of modeling nonlinear concept trajectories is interesting, and the qualitative and quantitative results are encouraging.

Critical weaknesses

My primary concern is the relationship to Wurgaft et al. (2026), Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior. Both works explore nonlinear concept geometry and steering along curved trajectories in activation space.

Although this paper introduces a KAN-based parameterization and focuses on text-to-image models, I would encourage the authors to further clarify the conceptual and methodological distinctions from this closely related work. In particular, it would be helpful to discuss what additional insights or capabilities the proposed formulation provides beyond existing manifold steering approaches.

Questions for authors

Necessity of KANs. Could the authors compare KANSteer with simpler nonlinear trajectory parameterizations, such as cubic splines or small MLPs? This would help clarify whether the improvements come from the KAN architecture or nonlinear steering more generally.

Geometry of concept representations. The method assumes that ordered visual attributes can be described by a one-dimensional trajectory. How well does this assumption generalize to more complex concepts involving multiple interacting attributes?

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

This paper documents KANSteer, an approach to modify text-to-image model hidden activations towards a target representation along a non-linear trajectory. They achieve this by prompting an image generating transformer diffusion model with a variation of a single attribute, fitting non-linear functions using a Kolmogorov-Arnold Network (KAN) encoder-decoder as the attribute direction estimator, and computing and applying the steering vector at inference time. The authors demonstrate better steering adherence than a linear steering approach, and better extrapolation on several image generation metrics on the FLUX.1 image generation model.

Critical strengths

This is a novel technique which combines two disparate ideas: KANs and image-generation model mechanistic interpretability, with a reasonable justification.

The extrapolation analysis provides good evidence that the function being learned is mapping to a real semantic trajectory. This is important to demonstrate when fitting non-linear mappings to neural networks, because it is easier to overfit than a linear mapping. The qualitative results do seem to show better steer-ability than the linear model.

They clearly establish the limitations of the linear representation hypothesis for traversing the latent space in text-to-image models, and engage with the relevant literature on this subject.

The paper is well worded and organized, with easily understood figures and captions. The PCA projections make it clear to see the limitations of the linear approach.

Critical weaknesses

While the results do show improvement over a linear steering, I don't believe this was ever in question, and the results are unsurprising. Adding more parameters to the steering function would expectedly improve the steering of in-distribution data. Most of the metrics used measure output quality in relation to the prompts, but KANSteer is explicitly designed to better fit these multiple prompts. The more interesting unresolved question is if the KAN approach is better than other non-linear steering methods, and why. That evaluation is missing, and is definitely necessary. The motivation to use KANs is not entirely clear to me, "they preserve an explicit scalar coordinate while allowing the trajectory to be nonlinear", but why not some other non-linear learned function like an MLP? I think this deserves a deeper motivation and analysis against other non-linear methods.

The other main concern is regarding overparameterization of KANSteer. "The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability?" by Sutter et al 2025 demonstrates that it is possible to fit any network (including randomly parameterized networks) to any algorithm using a sufficiently expressive non-linear mapping. So this concern must be addressed, perhaps by ablating the number of parameters in the KAN. How small can the KAN be before it loses its ability to represent the curve? That is the more interesting question than "can we steer the model along a semantic direction". I believe a demonstration that the approach works with a minimal number of parameters on two model families would bolster your argument. Also a negative control on a random network would establish whether the observed nonlinear structure actually depends on the learned representations, or can be spuriously fit with an sufficiently expressive function.

The claim on interpretability is strongly in question. While it is clear that the non-linear steering works better, that does not mean it is more explainable. In fact it might be less interpretable than the weaker, linear steering. I would remove any claims that KANSteer provides an interpretable explanation for the non-linear representation. Without more general demonstration of extrapolation, I don't think the provided evidence points to any interpretable explanation to the model's representation.

None of these weaknesses are insurmountable, so long as the results hold across multiple models and number of parameters.

Questions for authors

Did you evaluate your method against other non-linear steering approaches? What were those results?

Did you evaluate KANSteer on another family of models, or perhaps a randomly parameterized model? Do the positive results still hold for both?

Did you ablate the size of KANSteer? What was the process of hyperparameter selection?

Reviewer_j7rMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper shows how many visual concepts require a non-linear steering curve to transition between different attributes, e.g. straight to curly hair. They then propose a KAN-based method to parameterize such non-linear steering and both quantitatively and qualitatively show how this clearly outperforms linear steering.

Critical strengths

The formalization such as KAN and the loss terms are sensible without being overly complicated, to my understanding.

Te qualitative results are quite impressive! Linear steering’s failure modes can be intuitively seen. Also, Sec 5.3 is a nice natural follow-up question for any non-linear curve since it is harder to extrapolate curves compared to lines.

Critical weaknesses

At many points there is some missing details or clarity: Sec3 does not really explain how this learned steering is actually learned (eg L196 just states it). If it is just difference of means for both poles, why not state it? Also, not defined what “delta activation” is here. In general it feels like a lot of details from later sections should have been introduced here? With the actual “quantitative results” only referenced but no details beyond that mentioned. The appendix A also references one equation (presumably from Sec3) that never actually exists. While, most people are roughly aware how KANs are different from standard MLPs, a short formal reminder would have been useful! Also, it is never really stated clearly how many concepts are steered (and what their attributes are, how they are chosen, etc)

I would like to see more baselines for how to model such a non-linear trajectory? A MLP/polynomial curve or a simpler baseline from Goodfire’s paper(s) like centroid-spline fitting. Especially since the related work mentions how the novelty here lies in the parametric approach of using the KAN, it would be nice to compare to these other ones directly then.

Questions for authors

While I like the KAN approach here and think they are a natural fit, could you run simpler or more established baselines?

Comments on the paper

2 comments

Thank you for the updated manuscript. I've taken a look at the revision.

W1 is addressed, and you present strong results. You include new ablations over MLP, evidence of superior generalization, and reasonable argument for KANSteer over an MLP. My concern here is addressed.

W2/Q3.1 is also resolved. This is sufficient evidence to me to suggest it is not overparameterized. The random-network negative control is not included, but you show a practical benefit here and constrain the possibility of overfitting.

Q2 is directly addressed with the SANA experiments, comprehensively analyzing the baseline comparisons, extrapolation, and component ablations.

Q3.2 is also addressed. The grid search detail is included in the appendix.

W3 is mostly addressed. You present good evidence that the system effectively steers generated attributes, and that it captures regularities in the activation geometry. The interpretation of this as a mechanistic understanding or a simply useful nonlinear coordinate system is still open, but the claims have been narrowed to address this.

I am updating my review to a weak acceptance. Thank you for the great read!

Dear Reviewer,

Thank you for raising your score. We are glad the ablations, the SANA experiments and the added details addressed your concerns. Thank you again for your time and constructive feedback, which really helped us in improving our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.