acceptodds
Under review as a conference paper at ICLR 2027

LOFT: Latent Optimal Transport-guided Feature Transformation for Robust Automatic Modulation Recognition

Abstract

Automatic Modulation Recognition (AMR) seeks to infer the modulation schemes of received signals without prior knowledge. While deep learning has greatly advanced AMR, severe noise in low signal-to-noise ratio (SNR) regimes corrupts learned feature representations, thereby degrading recognition performance. Unlike previous methods that focus on data augmentation or signal denoising, this paper investigates low-SNR AMR failures through the lens of shortcut learning. As SNR decreases, models tend to take shortcuts by collapsing their predictions onto a few dominant classes, leading to class confusion. To tackle this issue, we propose **L**atent **O**ptimal Transport-guided **F**eature **T**ransformation (**LOFT**), a plug-and-play framework for reversing feature collapse. Motivated by the semantic invariance of modulation classes across SNRs, LOFT constructs clean class prototypes as semantic anchors and formulates latent recovery through an optimal transport-guided alignment objective. A lightweight transport adapter is then trained to predict feature displacements that move degraded representations toward clean semantic anchors in a single step. Furthermore, LOFT utilizes confidence-aware routing to selectively activate latent correction without requiring SNR information during inference. Extensive experiments on benchmark datasets demonstrate that LOFT consistently improves AMR performance, yielding an average accuracy gain of 1.63% across all SNR conditions and a 3.04% improvement in severe noise conditions. Code will be released upon acceptance.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
automatic modulation recognition, cognitive radio, optimal transport, adapter

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RZvYOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LOFT, a lightweight framework for improving low-SNR automatic modulation recognition while keeping the pretrained model frozen. It trains a residual feature adapter using an optimal-transport-inspired prototype alignment loss and a classification loss, with high-SNR class prototypes serving as reference targets and batch-level prediction entropy determining the correction state during inference. Experiments on two RadioML datasets across six backbones report improved recognition accuracy with modest additional parameters and training cost.

Critical strengths
  1. Consistent accuracy gains across six backbones on two RadioML datasets support the method’s applicability across different AMR architectures.
  2. The lightweight adapter requires only three training epochs while keeping the pretrained model frozen, offering an efficient way to improve existing models.
  3. Viewing low-SNR recognition as a latent feature recovery problem offers a fresh perspective and enables targeted improvements in noise robustness.
Critical weaknesses
  1. The theoretical analysis controls a Wasserstein upper bound through class-mean regression, but does not establish that optimization brings corrected features closer to the full clean distribution.
  2. Evaluation is limited to datasets with relatively few modulation classes, leaving performance on richer modulation sets, such as RML2018.01A and HisarMod2019.1, unclear.
Questions for authors
  1. Does LOFT empirically reduce the Wasserstein distance to the clean feature distribution, and how large is the residual dispersion term in the theoretical bound?
  2. How does LOFT perform on datasets with more diverse modulation classes, such as RML2018.01A and HisarMod2019.1?
1 comment
AuthorsOpenReview ↗

Response to Reviewer RZvY: Theoretical Validation and Generalization to Diverse Modulation Benchmarks

Thank you for your thoughtful feedback! We appreciate your positive assessment of our novelty, writing quality, efficiency, and generality. We have carefully addressed the concerns raised.

[W1&Q1] Quantitative Analysis of Theoretical Propositions.

We empirically verify that LOFT reduces the class-conditional Wasserstein discrepancy to clean reference features by 28.55%. We also quantify the residual dispersion term in Theorem 1 and demonstrate a 23.90% reduction in the theoretical upper bound.

1. Wasserstein distance reduction. We compute empirical W2W_2 using exact discrete OT (unregularized EMD with squared Euclidean cost) [1] on RML2016.10A. Each evaluation uses 128 low-SNR test features and 128 high-SNR (≥10\geq 10 dB) clean reference features.

MethodMean Empirical W2W_2 ↓Abs. ChangeRel. Change
Backbone AWN35.40--
+ LOFT25.29-10.11-28.55%

LOFT reduces the mean empirical W2W_2 by 28.55%, improving 10/11 classes. The sole exception is AM-SSB, consistent with the shortcut-dominance phenomenon illustrated in Fig. 1: the frozen backbone tends to misclassify samples from other modulation classes as AM-SSB, causing their representations to collapse toward the AM-SSB feature region. By mitigating this prediction bias, LOFT redistributes the collapsed representations toward their correct class-conditional clean distributions, improving overall alignment.

2. Residual dispersion and Theorem 1 upper bound. From Theorem 1 (Eq. 24),

\leq \sqrt{\mathbb{E}\||z+v_\phi(z,0)-\mu_y\||_2^2} + \sqrt{\operatorname{Tr}(\Sigma_y)}.$$ The first term measures the adapter-dependent discrepancy from the clean class prototype, while the second quantifies intrinsic clean-feature dispersion. We estimate both terms using 4,000 high-SNR clean training features per class. | Bound Component | Backbone | + LOFT | Rel. Change | |:---|---:|---:|---:| | Adapter-dependent term | 33.93 | **23.24** | **-31.51%** | | Residual $\sqrt{\operatorname{Tr}(\Sigma_y)}$ | 10.81 | 10.81 | - | | **Total Upper Bound** | **44.74** | **34.05** | **-23.90%** | The residual term is computed as $\sqrt{\operatorname{Tr}(\Sigma_y)}=\sqrt{\frac{1}{N_y}\sum_i\|z_i^{\mathrm{clean}}-\mu_y\|_2^2}$. Its class-averaged value is **10.81**. It remains unchanged because both methods share the same clean reference distribution. The adapter-dependent term contributes **75.8%** of the original upper bound and decreases by **31.51%** after LOFT, tightening the mean bound by **23.90%**. The bound reduction (**10.69**) is comparable to the empirical $W_2$ decrease (**10.11**), providing consistent theoretical and empirical evidence of improved feature alignment. Together, these findings support the theoretical interpretation of LOFT as a feature transformation that reduces the distributional discrepancy. > **[W2&Q2] Generalization to More Diverse Datasets.** To address this concern, we conducted six additional evaluations on the RML2018.01A [2] and HisarMod2019.1 [3] datasets, covering 24 and 26 modulation classes, respectively, and using 1024-length sequences. **1. Consistent improvements.** We evaluate LOFT on three architecturally distinct backbones (AWN, GRU2, and PETCGDNN). In all experiments, we keep the pretrained backbone frozen and train only the lightweight LOFT adapter for 5 epochs. | Backbone | RML2018.01A | Abs. Gain | HisarMod2019.1 | Abs. Gain | | :--- | :---: | :---: | :---: | :---: | | `AWN` (Base / **+LOFT**) | 61.86% / **62.35%** | **+0.49%** | 60.31% / **61.54%** | **+1.23%** | | `GRU2` (Base / **+LOFT**) | 63.72% / **64.27%** | **+0.55%** | 65.12% / **66.03%** | **+0.91%** | | `PETCGDNN` (Base / **+LOFT**) | 61.94% / **62.55%** | **+0.61%** | 59.81% / **60.76%** | **+0.95%** | LOFT improves accuracy in all six backbone–dataset combinations, with absolute gains of 0.49 to 1.23 percentage points. These results support LOFT's adaptability to larger class sets and 8× longer sequences with frozen backbones. **2. SNR-dependent gains.** On HisarMod2019.1, backbones reach 40% accuracy at −20 dB, indicating usable features; LOFT adds 0.91–1.23 pp overall. On RML2018.01A, −20 dB accuracy is near chance (4.17%), limiting recoverable information. Yet LOFT improves overall accuracy by 0.49–0.61 pp, suggesting benefits where discriminative features remain. **Revision:** We will update Section 5 and Appendix D.5 to present these results. **References:** [1] G. Peyré and M. Cuturi, “Computational Optimal Transport,” *Foundations and Trends in Machine Learning*, 2019. [2] T. J. O’Shea et al., “Over-the-Air Deep Learning Based Radio Signal Classification,” *IEEE JSTSP*, 2018. [3] K. Tekbıyık et al., “Robust and Fast Automatic Modulation Classification with CNN under Multipath Fading Channels,” in *VTC2020-Spring*, 2020.
Reviewer_WVnYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LOFT, a plug-and-play latent feature correction module for low-SNR automatic modulation recognition (AMR). The authors argue that low-SNR failures arise from shortcut learning that induces semantic representation collapse. LOFT constructs class prototypes from high-SNR samples as semantic anchors, trains a lightweight transport adapter to predict residual displacements toward these prototypes, and uses batch-level prediction entropy for selective activation of the correction. Experiments on RML2016.10A/B with several frozen backbones show modest accuracy gains, particularly at low SNR, with very low parameter and training overhead.

Critical strengths

The diagnosis of low-SNR AMR failure as prediction collapse onto a few dominant classes, supported by training dynamics and latent-space visualizations, is a useful perspective that goes beyond generic noise robustness arguments. The resulting design is genuinely lightweight (≈0.05 M parameters, 3 epochs) and leaves the backbone frozen, which is practically appealing for resource-constrained communication settings. The batch-level entropy routing is a reasonable attempt to avoid degrading already-reliable high-SNR representations without requiring explicit SNR estimation.

Critical weaknesses

The central claim that the work “formulates latent feature recovery as an optimal transport problem” is overstated. In practice the method reduces to prototype-directed residual regression toward the first moment μ_y − z. The appendix bound shows only that this regression controls the adapter-dependent term in a loose Wasserstein-2 upper bound to the empirical endpoint; the residual √Tr(Σ_y) term is uncontrolled, and no actual Monge map or entropic OT plan is computed. Without a direct comparison to discrete OT or Sinkhorn baselines, it is unclear how much performance is lost (or gained) by the prototype surrogate.

Aligning low-SNR features to high-SNR class prototypes with a lightweight residual adapter is already common in domain adaptation, test-time adaptation, and prototype-network literature. The paper does not articulate a clear, substantial distinction from these existing approaches beyond the specific AMR application and the binary transport-state abstraction.

The confidence-aware routing relies on batch-level entropy of the marginal prediction distribution and is only validated under an artificial streaming protocol that groups samples of similar SNR. Its behavior under single-sample inference or when a batch contains mixed SNRs is not examined, yet both settings are realistic.

Evaluation is limited to the two classic RML2016 datasets. No results are provided on datasets with longer sequences, larger numbers of modulation classes, or more recent/realistic channel conditions, leaving the generality of the observed gains unclear.

Questions for authors
  1. Why was a true discrete OT (or entropic/Sinkhorn OT) solver not included as a baseline against the prototype-regression adapter? What is the quantitative performance difference, and how does the computational cost compare?

  2. What are the concrete methodological distinctions between LOFT and existing prototype-alignment or residual-adapter techniques from domain adaptation and test-time adaptation? Which components are novel beyond the AMR setting?

  3. How does the confidence-aware routing behave under single-sample inference and under batches that mix different SNR levels? Are there failure cases, and can the authors provide quantitative results for these regimes?

  4. How does LOFT scale to datasets with substantially longer input sequences or a larger number of modulation classes? Are the reported gains preserved?

3 comments
AuthorsOpenReview ↗

Response to Reviewer WVnY: Comparison with OT Solvers (Part 1)

We thank you for your constructive comments and appreciate your positive remarks regarding the innovative perspective and practical potential of LOFT! Below, we address your concerns regarding the OT formulation, clarify the differences from prior methods, discuss practical deployment, and further elaborate on our core motivation. We hope these additional experiments and clarifications will resolve your concerns and encourage you to reconsider your assessment.

[W1&Q1] Sinkhorn OT Solver vs. Prototype-Regression Surrogate.

LOFT optimizes an OT-inspired prototype-discrepancy surrogate, W2(qϕ,y,s,δμy)W_2(q_{\phi,y,s},\delta_{\mu_y}), rather than explicitly recovering a Monge transport map. Our original design prioritizes computationally efficient feature correction. Following your suggestion, we implemented an entropy-regularized Sinkhorn OT baseline [1] using the Python Optimal Transport (POT) library [2] under matched experimental settings. LOFT achieves 0.59 percentage points higher accuracy with a 6.62× speedup. Furthermore, empirical evaluations show that LOFT reduces the exact discrete W2W_2 distance [3] to the full clean distribution νy\nu_y by 28.55%, providing direct evidence of improved distributional alignment beyond prototype discrepancy alone.

1. Controlled Comparison: Sinkhorn OT vs. LOFT. Using the same frozen backbone, adapter architecture, optimizer, loss weights, and 3 training epochs on RML2016.10A, we replace LOFT's prototype displacement targets for low-SNR samples with Sinkhorn-derived barycentric displacements. Within each class-SNR group, we solve entropy-regularized discrete OT between degraded features and clean reference features (s≥10 dBs\geq10\text{ dB}) using squared Euclidean costs CC, POT's log-domain Sinkhorn solver, ε=0.05⋅median(Cij>0)\varepsilon=0.05\cdot\mathrm{median}(C_{ij}>0), and up to 2,000 iterations. The resulting coupling PP defines the displacement target: uiOT=∑jPijzjclean∑jPij−zi.u_i^{\mathrm{OT}}=\frac{\sum_jP_{ij}z_j^{\mathrm{clean}}}{\sum_jP_{ij}}-z_i. High-SNR samples retain the original prototype supervision. This design focuses the comparison on the choice of displacement supervision while retaining the same adapter architecture and training objective.

MethodTest Acc. (%) ↑Gain (pp)Feat. ExtractOT PrecomputeAdapter TrainTotal Time ↓
Frozen AWN61.41%-----
+ Sinkhorn OT62.32%+0.911.78 s51.47 s10.29 s63.55 s
+ LOFT (Ours)62.91%+1.501.78 s-7.82 s9.60 s

LOFT outperforms Sinkhorn OT by 0.59 pp while reducing total preparation and training time by 84.9% (9.60 s9.60\text{ s} vs. 63.55 s63.55\text{ s}). Because both methods share the same lightweight adapter, the overhead in Sinkhorn OT stems mainly from explicit transport-plan precomputation.

2. Why Can Prototype Supervision Be Effective Under Severe Noise? Under severe noise, local Sinkhorn couplings may be sensitive to distorted feature geometry and finite-sample fluctuations. Their barycentric targets depend on pairwise source–target costs, whereas LOFT uses class prototypes estimated from the full clean reference bank, avoiding reliance on individual noisy correspondences. This offers a plausible explanation for the observed performance difference. Our comparison shows that local entropic OT supervision does not improve overall recognition accuracy over prototype regression in this setting, despite substantially higher computational cost.

3. Empirical Validation of Theorem 1 and Residual Dispersion. To assess the practical relevance of the bound, we measure empirical W2W_2 via exact unregularized discrete OT [3] and quantify both bound components. LOFT reduces the mean W2W_2 by 28.55% (35.403→25.29435.403\to25.294). The fixed clean-dispersion term Tr⁡(Σy)\sqrt{\operatorname{Tr}(\Sigma_y)} averages 10.812, while the adapter-dependent prototype-discrepancy term accounts for 75.8% of the initial upper bound and decreases by 31.51% (33.930→23.23933.930\to23.239). Consequently, the mean upper bound decreases by 23.90% (44.742→34.05144.742\to34.051). Although the residual dispersion is not optimized by LOFT, it remains fixed under the shared clean reference distribution. These results provide empirical support for the surrogate without establishing bound tightness. Please refer to our response to Reviewer RZvY [W1&Q1] for the full experimental protocol and bound decomposition.

References:

[1] M. Cuturi, “Sinkhorn Distances: Lightspeed Computation of Optimal Transport,” Advances in Neural Information Processing Systems (NeurIPS), 2013.

[2] R. Flamary et al., “POT: Python Optimal Transport,” Journal of Machine Learning Research, 2021.

[3] G. Peyré and M. Cuturi, “Computational Optimal Transport,” Foundations and Trends in Machine Learning, 2019.

AuthorsOpenReview ↗

Response to Reviewer WVnY: Methodological Differences and Generalization Analysis (Part 2)

[W2&Q2] Distinction from Previous Works.

LOFT's central contribution is to reinterpret the continuous transport-time parameter in OT- and flow-matching-inspired modeling as discrete, physically meaningful degradation states derived from SNR, and to construct feature-displacement supervision directly from a frozen pretrained backbone. This differs from conventional prototype alignment, residual-adapter techniques, and end-to-end domain adaptation. To clarify these distinctions, we consider three representative prototype-alignment paradigms:

(1) Domain Adaptation: SBPA [1] combines adversarial domain alignment with softmax-weighted prototype alignment by jointly optimizing the feature extractor and adaptation modules, while AutoDenoise-AMR [2] employs two-stage prototype-guided denoising training. Unlike these methods, LOFT keeps the pretrained backbone and classifier frozen and directly constructs high-SNR class prototypes in their existing feature space to supervise degradation-dependent displacements. (2) Test-Time Adaptation: T3A [3] updates class prototypes using pseudo-labeled test samples selected by prediction confidence. However, this selection relies on prediction confidence being informative about correctness, which can be unreliable under the low-SNR feature collapse observed in our study. (3) Adapter-Based Adaptation: lusterAdapter [4] learns multiple class prototypes from few-shot labeled samples for adaptation; DR-Adapter [5] rectifies domain shifts through feature-statistic perturbation and cyclic correction; and EDDA [6] adopts a divide-and-conquer strategy for domain-dependent classification. In contrast, LOFT learns a degradation-conditioned feature transformation through a two-state vector field vϕ(z,t)v_\phi(z,t), enabling stronger correction of degraded features.

Quantitative comparisons with representative baselines are provided in our response to Reviewer kPiS [Q2].

[W3&Q3] Routing Behavior.

We evaluated LOFT under single-sample inference (B=1B=1) and mixed-SNR batches. (1) For single-sample inference (B=1B=1), naive sample-wise entropy fails due to spurious overconfidence under low-SNR shortcut collapse; however, maintaining a lightweight sliding FIFO queue (W=16W=16) substantially mitigates this degradation, improving accuracy from 47.56% to 55.52%. (2) For mixed-SNR batches, LOFT remains robust under uniform global mixing (59.27%59.27\%).

1. Quantitative Evaluation Across Inference Regimes. Using the frozen GRU2 backbone and pretrained LOFT adapter, we evaluate five inference protocols across the test set:

Inference ProtocolOverall Acc. (%)Gain vs. Base (pp)Low-SNR (<0 dB<0\text{ dB})High-SNR (≥0 dB\ge 0\text{ dB})
Frozen GRU258.86%—31.75%85.97%
Single-Sample47.56%-11.3028.42%66.70%
Single-Sample + FIFO (W=16W=16)55.52%-3.3429.36%81.68%
Global Mixed-SNR Batch59.27%+0.4131.91%86.63%
Streaming Inference60.77%+1.9134.97%86.57%

2. Failure-Case Analysis. When B=1B=1 without memory, H(B)\mathcal{H}(\mathcal{B}) reduces to individual predictive entropy. Because low-SNR shortcut collapse yields spuriously low sample entropy toward AM-SSB (Sec. 3.4), naive B=1B=1 routing misclassifies collapsed samples as reliable. Incorporating a sliding FIFO queue aggregates recent predictions to provide a more stable estimate of the marginal prediction distribution, partially mitigating the degradation observed under single-sample inference.

[W4&Q4] Generalization Analysis.

On RML2018.01A (24 classes, L=1024L=1024) and HisarMod2019.1 (26 classes, L=1024L=1024), LOFT consistently improves all tested frozen backbones: AWN gains +0.49% / +1.23%, GRU2 gains +0.55% / +0.91%, and PETCGDNN gains +0.61% / +0.95%. Please refer to our response to Reviewer RZvY [W2&Q2] for the complete performance table and sequence-length analysis, which has also been added to Appendix D.5.

References:

[1] J. Li et al., “Unsupervised Domain Adaptation via Softmax-Based Prototype Construction and Adaptation,” Information Sciences, 2022.

[2] J. Ma et al., “AutoDenoise-AMR: Prototype-Guided Denoising Training Framework for Automatic Modulation Recognition,” ICASSP, 2025.

[3] Y. Iwasawa et al., “Test-Time Classifier Adjustment Module for Model-Agnostic Domain Generalization,” NeurIPS, 2021.

[4] M.-H. Guo et al., “Tuning Vision-Language Models With Multiple Prototypes Clustering,” TPAMI, 2024.

[5] J. Su et al., “Domain-Rectifying Adapter for Cross-Domain Few-Shot Segmentation,” CVPR, 2024.

[6] X. Zhang et al., “EDDA: An Efficient Divide-and-Conquer Domain Adapter for Automatics Modulation Recognition,” IEEE JSTSP, 2025.

Dear Reviewer WVnY,

Thank you for your careful review and constructive comments on our paper. In response to your initial review, we have posted point-by-point replies addressing all the issues you raised.

We would appreciate it if you could take a look at our responses and let us know whether any questions or concerns remain. We are happy to provide further clarification or discussion at your convenience. Thank you again for the time and effort you have devoted to our submission!

Best regards, Authors

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

This paper proposes a lightweight feature transformation framework designed to for AMR tasks under low SNR.

Critical strengths

The authors attribute low-SNR classification failures to shortcut learning and semantic representation collapse, where noisy signals from different modulation classes become entangled in the latent space and predictions concentrate on a few dominant classes. To address this, LOFT constructs class-specific prototypes from high-SNR features and introduces an optimal transport motivated displacement objective to guide degraded representations towards clean semantic anchors. A lightweight residual adapter learns this transformation, while a batch-level entropy-based routing mechanism determines whether correction is required during inference without explicit SNR information. Experiments on RML2016.10A and RML2016.10B using 6 AMR backbones report improvements in recognition accuracy, particularly under severe noise, with relatively few additional parameters and a short adapter training period.

Critical weaknesses
  1. The novelty of the proposed optimal transport formulation is questionable, particularly given its close relationship to conventional prototype-based feature alignment. The paper claims to be the first to formulate latent feature recovery for low-SNR AMR as an optimal transport problem. However, the actual implementation does not solve the Monge mapping problem introduced in Equation (4).
  2. The paper does not sufficiently distinguish its contributions from existing domain adaptation and parameter-efficient AMR methods. However, the experiments only compare pretrained AMR backbones before and after applying LOFT. A more proper way should include comparison with SOTA modular adaptation and representation alignment methods.
  3. More importantly, the interpretation of low-SNR prediction collapse as shortcut learning is not sufficiently justified. the evidence in Figures 1 and 4 mainly shows that predictions become concentrated in certain classes and that latent representations exhibit poor separability.

Some minor issues:

  1. The use of strong expressions such as “catastrophic semantic collapse”, “profound geometric reshaping”, “fundamentally reconstructing”, and “definitively proves” are unnecessary and not properly justified.
  2. The parameter counts in Table 3 should be reconciled with the statements about a fixed 0.05M-parameter adapter. Training time should be reported with a clearly defined measurement scope.
  3. The heading “Details of the Implementation Details” in Appendix B.3 is redundant and should be revised to “Implementation Details”.
Questions for authors

Please see the weaknesses above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer kPiS: Method Modeling and Baseline Comparisons (Part 1)

Thank you for your insightful comments! We appreciate your positive remarks regarding the empirical strength and potential of our paradigm. Below, we address the concerns raised about method modeling and baseline comparisons in detail. We hope these clarifications will resolve your concerns and encourage you to reconsider your rating.

[W1] Clarification of the OT Formulation.

1. OT Motivation and Prototype Approximation. Equation (4) establishes the ideal distributional recovery objective, which motivates LOFT's prototype-directed surrogate. As explained in Section 3.2, explicitly estimating the Monge map requires sample-wise transport estimation. We therefore approximate the clean endpoint distribution by its first moment, yielding uyproto(z)=μy−zu_y^{\mathrm{proto}}(z)=\mu_y-z. Importantly, Appendix A.3 (Theorem 1) establishes a Wasserstein upper bound:

\leq \sqrt{\mathbb{E}\||z+v_\phi(z,0)-\mu_y\||_2^2} + \sqrt{\operatorname{Tr}(\Sigma_y)}.$$ Thus, prototype regression controls the adapter-dependent component of the distributional discrepancy to the full clean endpoint. We acknowledge that the residual dispersion is not optimized, **as clarified in Appendix A.4.** **2. Distinction from Prototype Alignment.** LOFT learns a deterministic, degradation-conditioned discrete vector field $v_\phi(z,t)$, motivated by OT-based distributional recovery formulation. Prototype-directed displacement serves as an efficient surrogate, theoretically justified by a Wasserstein upper bound. **3. Quantitative results.** We further validate this design empirically. Compared with an explicitly implemented Sinkhorn OT baseline under matched settings, LOFT achieves 62.91% accuracy, while reducing total preparation and training time from 63.55s to 9.60s. Moreover, exact discrete OT evaluation shows that LOFT reduces empirical $W_2$ to the clean distribution by 28.55%. The theoretical upper bound also decreases by 23.90%, while the residual dispersion remains fixed at 10.812. These results support the effectiveness of the OT-inspired prototype surrogate as a computationally efficient approximation. Please refer to our responses to **Reviewer WVnY [W1&Q1] and Reviewer RZvY [W1&Q1]** for experimental details. **We will further clarify this distinction in the revised manuscript.** > **[W2] Comparison with Representative Methods.** To address this concern, we conducted additional controlled experiments, including conventional prototype alignment (DA-Proto), residual feature adaptation (DA-Residual), and five representative adaptation methods: SBPA [1], T3A [2], lusterAdapter [3], DR-Adapter [4], and EDDA [5]. For a consistent comparison, all methods use the same pretrained backbone, data split, and 3-epoch adapter-training protocol. The results are summarized below. | Method | Mechanism under Frozen Backbone | Accuracy (%) | $\Delta$ Base (pp) | | :--- | :--- | :---: | :---: | | Frozen `GRU2` | - | 58.86% | - | | `DA-Proto` | Prototype alignment | 58.71% | -0.15 | | `DA-Residual` | Residual adaptation | 58.80% | -0.06 | | `SBPA` [1] | Softmax prototypes + domain alignment | 58.13% | -0.74 | | `T3A` [2] | Test-time prototype adjustment | 49.08% | -9.78 | | `lusterAdapter` [3] | Multi-prototype residual adapter | 58.70% | -0.17 | | `DR-Adapter` [4] | Statistical perturbation + cyclic rectification | 58.78% | -0.08 | | `EDDA` [5] | Divide-and-conquer adapter | 59.33% | +0.47 | | **`+ LOFT (Ours)`** | **Deterministic displacement field** | **60.77%** | **+1.91** | Existing prototype alignment and residual-adapter baselines yield slightly lower accuracy than the frozen backbone (−0.06 to −0.74), while T3A exhibits a substantial drop (−9.78). This degradation may be related to unreliable pseudo-labels under the low-SNR feature collapse observed in our study. In contrast, EDDA gains only +0.47, while LOFT achieves a remarkable +1.91 with just 3 epochs of adapter training. These results support the effectiveness of LOFT's degradation-conditioned feature correction under the evaluated protocol. For a detailed discussion of LOFT's distinctions from domain adaptation, test-time adaptation, and adapter-based approaches, please refer to our response to **Reviewer WVnY [W2&Q2]**. **References:** [1] J. Li et al., “Unsupervised Domain Adaptation via Softmax-Based Prototype Construction and Adaptation,” *Information Sciences*, 2022. [2] Y. Iwasawa et al., “Test-Time Classifier Adjustment Module for Model-Agnostic Domain Generalization,” *NeurIPS*, 2021. [3] M.-H. Guo et al., “Tuning Vision-Language Models With Multiple Prototypes Clustering,” *TPAMI*, 2024. [4] J. Su et al., “Domain-Rectifying Adapter for Cross-Domain Few-Shot Segmentation,” *CVPR*, 2024. [5] X. Zhang et al., “EDDA: An Efficient Divide-and-Conquer Domain Adapter for Automatics Modulation Recognition,” *IEEE JSTSP*, 2025.
AuthorsOpenReview ↗

Response to Reviewer kPiS: Evidence Supplementation and Minor Revisions (Part 2)

[W3] Further Evidence and Analysis.

1. Original Motivation. Shortcut learning refers to a model relying on easily accessible but non-robust cues [1]. Under severe noise, fine-grained modulation-specific features become increasingly corrupted, making reliable class discrimination difficult. As illustrated in Figures 1 and 4, the frozen backbone exhibits a pronounced bias toward AM-SSB, rather than uniform or random prediction degradation. We argue that this behavior reflects a shortcut-like reliance on non-robust representations. This diagnosis is also consistent with the positive assessments from Reviewers WVnY and RZvY.

2. Asymmetric Representation Drift. As discussed in Section 3.1, at -20 dB, AM-SSB exhibits nearly zero semantic drift (0.0073) from its clean prototype, whereas the other 10 classes undergo substantial representation degradation. This pronounced asymmetry suggests that the AM-SSB representation remains relatively stable under severe noise, potentially making it a dominant attractor for degraded representations from other classes.

3. Additional Quantitative Validation. To further substantiate this interpretation, we evaluate the frozen AWN from three complementary perspectives: prediction distributions, prototype distances, and classifier decision margins. Specifically, we measure (i) the fraction of non-AM-SSB samples closer to the clean AM-SSB prototype than to their respective class prototypes, (ii) the ratio of mean squared distances to the corresponding and AM-SSB prototypes, and (iii) the fraction with positive AM-SSB-vs.-true logit margins.

SNR (dB)Acc. (%)Dominant ClassDom. Pred. (%)Closer to AM-SSB Proto. (%)Dist. Ratio (Own/AM-SSB)Positive Logit Margin (%)
-209.77AM-SSB90.73100.0064.96×96.65
-1610.55AM-SSB84.09100.0055.16×92.30
-1217.64AM-SSB65.0598.8018.51×71.50
-1028.18AM-SSB47.3290.809.09×49.95
-839.41AM-SSB26.3684.254.46×27.85
-653.86QAM6416.8671.852.01×4.75
087.27AM-DSB14.360.900.33×0.65
1090.95AM-DSB14.500.700.25×0.45
  • Systematic Geometric Attraction: At −20 dB-20\text{ dB}, all non-AM-SSB samples are closer to the clean AM-SSB prototype than to their respective class prototypes. Their mean squared distance to their own prototypes is 64.96 times that to the AM-SSB prototype.
  • Classifier Decision Bias: At −20 dB-20\text{ dB}, 96.65% of non-AM-SSB samples exhibit positive AM-SSB-vs.-true logit margins, consistent with the 90.73% dominant-class prediction ratio across all test samples. Both geometric attraction and prediction bias progressively diminish as SNR increases.

Overall, these findings reveal systematic, SNR-dependent dominant-class bias across representation geometry, classifier decisions, and prediction distributions, going beyond mere feature inseparability. They provide strong empirical support for our interpretation of low-SNR prediction collapse as shortcut-like behavior. We will incorporate these analyses into the revised manuscript.

[Minor Issues]

1 & 3. Tone Moderation and Heading Fix. We have moderated strong expressions across the manuscript and corrected the Appendix B.3 heading to "Implementation Details".

2. Parameter Counts and Training Time Scope.

  • Reconciling Parameter Counts: LOFT operates at the latent level, so the adapter size scales with the frozen backbone's latent dimension DD. For backbones with D=128D=128 (e.g., GRU2, FEA-T, PETCGDNN), the adapter introduces 49,728 parameters (≈0.05 M\approx 0.05\text{ M}). For architectures with smaller latent dimensions, such as IQFormer (D=64D=64) and TransGroupNet (D=96D=96), the adapter requires only 0.02M–0.03M parameters. We have clarified this concept in the revised version.
  • Training Time Measurement Scope: Measured on a single RTX 4090 GPU (batch size 512, 3 epochs) on pre-extracted features, adapter training takes 7.82 s7.82\text{ s}. We have made this measurement scope in Section 5.

References:

[1] R. Geirhos et al., “Shortcut Learning in Deep Neural Networks,” Nature Machine Intelligence, 2020.

Dear Reviewer kPiS,

We sincerely appreciate your thoughtful feedback throughout the review process. Following your initial comments, we have submitted detailed responses to all the points raised.

We would be grateful if you could review our responses and let us know if you have any further questions or concerns. We would be more than happy to address them. Thank you again for your time and support!

Best regards, Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We thank the AC and reviewers for their thorough and constructive reviews. We are encouraged that the reviewers highlighted our novel perspective (Reviewers WVnY, RZvY), clear presentation (Reviewers kPiS, RZvY), consistent performance gains (Reviewers kPiS, WVnY, RZvY), and practical applicability (Reviewers kPiS, WVnY, RZvY). This feedback has been invaluable in helping us improve the manuscript.

In the revised version of the paper, we have added the following new sections that address the common questions and weaknesses raised by the reviewers; all revisions are highlighted in blue\textcolor{blue}{\textbf{blue}}:

Major revisions

  • Added quantitative analysis of theoretical propositions in Section 5 (Q3), including empirical Wasserstein distance measurements and a decomposition of the theoretical upper bound. (Reviewers WVnY, RZvY)
  • Added Sinkhorn OT Solver baseline in Table 2 to compare LOFT with true OT Solver in terms of recognition accuracy and computational cost. (Reviewers kPiS, WVnY)
  • Compared representation alignment methods in Appendix D.2, including prototype alignment, residual adaptation, test-time adaptation, and domain-specific adapters. (Reviewers kPiS, WVnY)
  • Generalized to RML2018.01A and HisarMod2019.1 in Appendix D.5 to assess generalization to richer modulation sets and longer observations. (Reviewers WVnY, RZvY)
  • Added routing robustness analysis in Appendix D.6 under various realistic settings. (Reviewers WVnY)
  • Included quantitative evidence of shortcut behavior in Section. 5 (Q1) and Appendix D.1. (Reviewer kPiS)

Minor revisions

  • Clarified the parameter quantities and training time of LOFT, and explained this in Section 5 (Q4). (Reviewer kPiS)
  • Clarified the OT-inspired nature of LOFT, emphasizing its prototype-directed displacement surrogate. (Reviewers kPiS, WVnY)
  • Expanded the related-work discussions in Section 2 and Appendix C.2 to distinguish LOFT from existing methods better. (Reviewers kPiS, WVnY)
  • Moderated some overly strong expressions and corrected the Appendix B.3 heading to "Implementation Details". (Reviewer kPiS)

We hope these revisions address all concerns and further strengthen the paper.

Thank you again for your time and helpful feedback.

Best regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.