acceptodds
Under review as a conference paper at ICLR 2027

Shortcut Lock-In: The Dynamics of Learning and Escaping Spurious Features

Abstract

Neural networks can improve at the core task while continuing to follow a spurious cue on conflicting inputs. When does further training allow useful task information to take control? We investigate this gap between capability and decision control through an evidence-transport account that separates the supply of corrective examples from the decision progress induced by their updates. A tractable two-route model shows how continued core learning during shortcut dominance leads to eventual escape under persistent conflict evidence, with distinct capture and escape timescales. During recovery, ViT-Tiny preserves coarse semantics while its sample geometry changes more than ResNet-18’s. Early measurements also carry timing information: on four prospectively held-out ViT-Tiny runs, first-ten-update estimates predict escape onset with 13.33% median absolute percentage error. Visual and multimodal experiments reproduce capture and recovery, and a five-seed GQA study extends the analysis to natural image–question pairs. These results distinguish the acquisition of useful task information from the training dynamics that make it decisive.

Primary area:
learning theory
Keywords:
shortcut learning, spurious correlations, training dynamics, evidence transport, representation dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DFWHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Spurious correlations is one of my most favorite and simple topics that can be easily understood. This paper tries to understand when in the presence of spurious features/cues, the model starts to escape from them and how and why does this escape vary across architectures.

Critical strengths

None. This is one of the most poorly written papers I have ever read - both technically and grammatically.

Critical weaknesses
  1. This is one of the most extremely poorly written paper that I have ever read on spurious correlations from the abstract till the Conclusion. Almost all the sections in this paper have technical issues in them. I put below several such examples of them directly from the paper (I rarely say this but it seems like to me that the authors have used AI tools for writing the whole paper without even proof-reading what they have written even though they say they have proof-read all the AI assisted content in their AI Use Statement section):
  • Fig 1 says that in early checkpoints, shortcut usage dominates. However, in Fig 2 for all 3 models, we can clearly see that in early checkpoints both shortcut (orange line) and core usage (blue line) are almost the same. This is completely conflicting. Moreover, Fig 1 is only reported for 1 seed even though the authors say they ran 5 seeds. I don't see any figures showing other seeds results.
  • Line 104 is so convoluted to even understand the digit-color task considered in this study.
  • There is no clarity regarding how P_align, P_swap, A(t) and C(t) are calculated in the main paper when they are originally introduced in line 129. Appendix B.1 tries to explain them but they are still never defined and explained how to even compute them. This is a major technical limitation given the authors base their paper's conclusions on these terms.
  • Line 136 says "main reliability sweep" but never even defines what it actually means. Moreover, in the same paragraph, the observation that they make by saying as pp increases (I believe pp here refers to the probability spurious feature appears in train set - also very difficult to understand in one-go), spurious features reliance increases, is a well-known fact from several prior works [1,2] and nothing is new about it.
  • In line 148, the authors say - "Reliability changes the opportunity to recover". This line does not even make any sense. I don't even think it is even a complete sentence, given the context in which they use the word "reliability".
  • In line 161, the authors introduce a new term called "Shallow-MM" but never defined it beforehand. I still don't understand what they even mean by it.
  • I don't even understand what Section 3 even tries to tell for the paper. For instance, what are: "logit allegiances", "two route model", "shortcut and core product margins", "U", "V", as defined in lines 216, 247-249. There many more such technical terms which are never defined in Section 3. Moreover, all the theoretical equations read like they are just forcefully added to the section to make it even more convoluted.
  • CKA similarity experiment discussed in Section 4 never explains why Tiny-ViT experiences later escaping from spurious cues compared to U-Net and ResNet18. It has never been casually explained in the paper.
  • Section 5 says that it is extended spurious cues to other modalities but the tasks for the new modalities are also never explained properly. I can't understand anything what the authors are even trying to say from this section.
  • Discussion and Conclusion sections read like broken sentences without ever summarizing the core contributions of the paper.
  1. Apart from the above mentioned technical issues in the paper (they are many more - I have listed as much as I can within my timeframe), the paper fails to even consider widely used spurious correlations dataset called - WaterBirds, CelebA, MNLI, HardImageNet. The authors should instead consider these datasets before even jumping to other new modalities in generative tasks.

  2. The authors say that Tiny-ViT observe late onset of escape compared to ResNet18 and U-Net. However, this is just an incomplete analysis to make any claim. The results should be verified for atleast 4-5 transformers (vision and language both) and 4-5 non-transformers models like ResNet, VGGNet, U-Net, DensNet. Also all transformer and non-transformer set of models should consist of multiple sizes to see if the trend is fixated to a particular size or any particular trend emerges among model sizes.

[1] Last layer re-training is sufficient for robustness to spurious correlations.

[2] Spurious Correlations in Machine Learning: A Survey

Questions for authors

Please see Weaknesses

Overall I believe the paper needs a major revision from basic thinking, technical, and writing perspectives before it can be submitted to any conference in the future, and I don't think any form of revision will help them improve their chances for the current conference submission. I hope the authors can significantly improve their paper based on the feedback provided. In the current stage I am not willing to accept it. (PS: I have not used any AI tool to review your paper).

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

The paper studies the gap between task-relevant information available to a network and the information that actually controls its decisions, examining how this gap evolves during training. The authors distinguish shortcut capture, crossover toward core-feature use, and successful escape. Controlled digit-segmentation experiments separate shape, which determines the target, from color, which provides a potentially misleading shortcut. The paper combines comparisons across architectures, a simplified two-route model, representation analyses, and early predictions of recovery timing.

Critical strengths

The paper clearly distinguishes the availability of useful core information from its influence on decisions when core and shortcut cues conflict.

The controlled experiments are a strength. Color swaps preserve digit identity while changing the misleading cue, enabling direct assessment of core-following and shortcut-following behavior.

The comparison of U-Net, ResNet-18, and ViT-Tiny reveals different recovery trajectories under common shortcut-reliability conditions. Tracking these trajectories provides information that final accuracy alone would miss.

The authors acknowledge important limitations, including finite observation horizons, differences between theoretical and empirical observables, depth–parameter-count confounding, and the distinction between initial escape and lasting recovery.

Critical weaknesses

The literature gap could be stated more precisely. Although the introduction discusses relevant studies, much of the detailed positioning appears in the appendix. The main text should clarify what existing theories already explain about recovery and which limitations prevent them from addressing the present setting.

In particular, the distinction from Izmailov et al. (2022) and other feature-learning studies deserves greater emphasis. The paper’s most convincing contribution is its empirical characterization of recovery timing across architectures. The proposed mechanistic account is potentially valuable, but the exposition should distinguish observed architectural differences from established causal explanations of those differences.

The forecasting procedure also requires more detail. Section 4.2 reports predictions from the first ten optimizer updates, but Appendix B.6 does not sufficiently specify the input measurements, prediction formula, calibration procedure, or configuration search. These details are necessary to reproduce and assess the forecasting results.

Questions for authors
  • How can the evidence-transport framework independently predict recovery time, given that Equation 5 uses an average transport rate defined over the interval ending at the unknown crossing time? Can this rate be estimated from early training measurements?
  • How accurately do these estimates predict recovery on held-out runs compared with simple baselines using architecture and shortcut reliability? Please also clarify how the predictor was calibrated and fixed before prospective evaluation.
  • My understanding is that the two-route model is not derived from full Transformer updates, as acknowledged in Appendix A.1. What evidence connects its mechanism to the recovery observed in ResNet and ViT? Can its state variables be related to measurable quantities in these networks?
Reviewer_Rx6kOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies a training setup where the data contains a core feature and an easier, spurious feature that acts as a shortcut. For a large fraction ρ of the examples the shortcut agrees with the label, and for the remaining fraction 1 − ρ it conflicts with it. With this setup, the paper studies whether and when the core feature takes control of the decision after the model has first learned the shortcut, across architectures (U-Net, ResNet-18, ViT-Tiny), and when the crossover between the shortcut and the core feature happens. The paper proposes an "evidence-transport" identity and a two-route toy model with separate capture and escape timescales, analyzes representation changes during recovery, reports that the crossover time can be predicted from the first ten updates, and extends the experiments to other shortcut carriers, a small multimodal task and GQA.

Critical strengths

The observation that, at the same shortcut reliability ρ, ViT-Tiny takes much longer to recover from the shortcut than U-Net and ResNet-18 is interesting.

Critical weaknesses
  1. The analysis does not explain the lag in recovery and is very preliminary. The two-route model of Section 3.2 does not account for the architecture, so it cannot explain why ViT-Tiny recovers more slowly than the CNNs. The shortcut-first behavior is also imposed through the scale κ\kappa (capture requires κ(2ρ−1)>1\kappa(2\rho - 1) > 1) rather than derived.

  2. The scope of the results is small and the experiments are tailored. The training details of the main sweep (optimizer, learning rates, batch size) are not given, the ViT-Tiny runs were extended depending on their trend, the carrier experiments use one seed per carrier, and in the GQA sweep depth and parameter count vary together.

  3. Simplicity bias is already known, and it is unclear what the empirical and theoretical evidence here adds.

  4. The paper requires a major rewriting. It introduces many terms (evidence transport, budget, focus, capture, crossover, escape, lock-in, availability–control gap) without enough support and lacks a coherent narrative. Most definitions needed to check the main-text claims appear only in the appendix.

  5. The notations and methods are not defined. The losses LaL_a and LcL_c (Section 3.1) are never defined, nor is the loss used in each experiment, or whether the training conflict distribution matches the swap distribution QswQ_{sw}. Several methods are not specified, such as the Shallow-MM task and its architecture, the evidence score and output fidelity of Figure 4c, and the end of the recovery interval used to align checkpoints.

  6. The predictive network of Figure 5 is not specified, and no details of the diagnosis are given. The paper does not say what is measured during the first ten updates, what model maps it to the crossover time, or how this model is fit. The 26-run results use a predictor whose configuration search included the four held-out runs, so only these four runs test it, and no baseline using only the architecture and ρ\rho is reported. The 13.33% error in the abstract therefore cannot be verified.

Questions for authors
  1. Could you define LaL_a and LcL_c, state the loss used in each experiment, and clarify whether the training conflict distribution matches QswQ_{sw}?

  2. Could you describe the predictor of Figure 5: which measurements are taken during the first ten updates, what model maps them to TcrossT_{cross}, and how it is fit? How does it compare with a baseline that uses only the architecture and ρ\rho?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.