acceptodds
Under review as a conference paper at ICLR 2027

Learning a Generative Inverse to Adapt the Forward: Source-Free Domain Adaptation for Sequence Regression

Abstract

Source-Free Domain Adaptation (SFDA) has emerged as a pragmatic paradigm for adapting models to practical scenarios without access to source data. Most existing SFDA methods focus on image classification and rely on entropy minimization or pseudo-labeling under the cluster assumption. Sequence regression instead has a continuous output space and does not provide discrete category codes. Simply extending classification-style methods to regression often fails to yield satisfactory results and may even lead to negative transfer. To address this issue, we propose Generative Inverse Regression Adaptation (GIRA), a general-purpose SFDA framework designed for sequence regression. Specifically, GIRA defines generalized classes over the continuous output space and uses a conditional generative model to learn the inverse mapping of the pre-trained model, generating associated target-style inputs. These resulting samples provide adaptation signals for updating the regression model on the target domain. Furthermore, uncertainty threshold filtering and feature scale constraints are integrated to mitigate negative transfer and ensure training stability. Across nine transfer scenarios in three tasks, GIRA reduces MAE by 20.87% on average relative to zero-shot and achieves the highest MAE improvement among the compared methods in most tasks.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Source-Free Domain Adaptation, Sequence Regression, Conditional Generative Model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CxptOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GIRA, a source-free domain adaptation framework for sequence regression. It uses generalized classes and a conditional generative model to construct adaptation signals from unlabeled target data. Experiments on three tasks and nine transfer scenarios show improvements over the zero-shot models and several adaptation baselines.

Critical strengths
  1. The paper addresses an problem in source-free domain adaptation for sequence regression. This is a meaningful research direction with potential practical value.

  2. The evaluation covers three different tasks and nine transfer scenarios, providing evidence that the approach can work across different regression settings. The ablation studies also offer some insight into the roles of the proposed components.

Critical weaknesses
  1. The main adaptation loss is applied to generated inputs rather than directly to real target inputs. The current model prediction also serves as the generator's conditioning input, and gradients propagate through this path. It is therefore unclear whether minimizing the loss improves predictions on real target data or mainly adjusts the conditioning and generated samples.

  2. The theoretical analysis relies on several assumptions, and the resulting risk bound depends on terms such as the generalized-class error and generated–target discrepancy, which are difficult to assess in the source-free setting. Although the appendix derives further bounds, it remains unclear how informative these bounds are under the assumptions required and how tightly they relate to the actual target risk.

Questions for authors
  1. How does conditioning the generator on the current prediction N(xT)N(x_T), rather than the fixed initial prediction yˉ(xT)\bar y(x_T), affect adaptation? Is there evidence that the gradient through the conditioning path is important?

  2. The GC-only ablation performs substantially worse than the zero-shot model on T#4. Which components are primarily responsible for preventing negative transfer, and how much does the generative inverse mapping contribute beyond regularization and filtering?

  3. GIRA performs less favorably in speech enhancement under larger domain shifts. How sensitive is the method to violations of the assumption that the pretrained model retains basic predictive capability on the target domain?

3 comments
AuthorsOpenReview ↗

Responses to Reviewer Cxpt

We thank the reviewer for the constructive review and for recognizing our research direction, evaluation and ablation studies. We respond to the concerns below.

C1. Does a loss on generated inputs improve real target predictions, or only adjust the conditioning?

We thank the reviewer for raising this point. Minimizing Eq. (12) improves the predictions on real target data, not only the conditioning or the generated samples. GG is frozen in Step 2, so NN is the only trained network, and every change of the condition or the generated samples results from a change of NN.

This works like self-supervised pseudo-labeling in classification SFDA (e.g., SHOT), where the correctly labeled majority of a class corrects its mislabeled minority. GG fills each generalized class, identified by the source prediction of one real input, with many target-style inputs paired with that identifier. Most pairs are correctly assigned, so fitting NN to them corrects its predictions on the target inputs they resemble, including the real ones (see gm9L-C1).

Unlike classification, where the classes are given, the GCs are defined by the model itself, so the position of an identifier can also be wrong. Conditioning GG on the current prediction N(xT)N(x_\mathcal{T}) lets the class position move with NN while the target yˉ(xT)\bar{y}(x_\mathcal{T}) stays fixed. Its gradient takes two paths. The shallow path, through the outer NN, corrects the wrongly assigned generated inputs. The deep path, through GG back to the inner N(xT)N(x_\mathcal{T}), moves the class position to where the generated population is consistent with its identifier. Since the inner N(xT)N(x_\mathcal{T}) is the prediction on the real input xTx_\mathcal{T}, this path adjusts the real-target predictions directly.

C2. How informative is the risk bound?

The bound is informative because its terms track the actual target risk. In its expanded form (Eq. 48), all data-dependent terms except the error–uncertainty envelope of Eq. (25) are label-free, namely LA\mathcal L_A, the generator Score, and the UTF threshold and coverage. The generator Score ranks the ten T#4 generator checkpoints exactly as the adapted MAE (Table 5). As the bound predicts, GIRA improves target accuracy once the generator is trained for 3,000 of 10,000 steps and the source model for 15 of 50 epochs (Tables 5, 17), i.e., when most generated pairs are correctly assigned (see Ru2P-C1). Like other domain adaptation bounds, it identifies which quantities control the risk rather than giving a numerical estimate.

C3. Does conditioning on N(xT)N(x_\mathcal{T}) rather than the fixed yˉ(xT)\bar{y}(x_\mathcal{T}), with its gradient path, matter?

We thank the reviewer for this question and ran this ablation on T#1, T#4 and T#5 (gm9L-C3, App. C.6). Conditioning on the fixed yˉ(xT)\bar{y}(x_\mathcal{T}) or stopping the gradient through N(xT)N(x_\mathcal{T}) still improves over zero-shot, while the full gradient path gives the best MAE. Both variants keep only the shallow path, which corrects the wrongly assigned generated inputs and provides most of the gain. The deep path adds the rest by correcting the class positions, i.e., the predictions N(xT)N(x_\mathcal{T}) on the real inputs (see C1), so the gradient through the condition is important.

C4. Why does GC-only fail on T#4, and what does the inverse contribute beyond regularization and filtering?

We appreciate this question. In Table 4, LSCC\mathcal L_{SCC} is what prevents negative transfer on T#4. Without it, MAE is 1592.41, close to GC-only (1603.67), whereas without LREG\mathcal L_{REG} (362.85) or UTF (313.12) the model still beats zero-shot (498.13). GC-only fails through a feature-scale drift that shifts the whole output by a constant offset (Fig. 3b). It is strongest on T#4, whose domains differ most (Fig. 4a).

The gain itself comes from the inverse mapping. With UTF, LREG\mathcal L_{REG} and LSCC\mathcal L_{SCC} kept identical, removing LGC\mathcal L_{GC} or replacing it by Gaussian-perturbation consistency or UTF self-training leaves MAE within 3% of zero-shot (see Ru2P-C2).

C5. How sensitive is GIRA to violating the basic-performance assumption?

App. C.4 (Table 17) varies source quality on T#4 with source checkpoints from epochs 10 to 50. At the weakest checkpoint, whose zero-shot MAE (910.24) is 83% higher than that of the final one, MAE worsens by only 1.63%. Gains are positive from epoch 15 (+11.68%) and grow to +47.96% at epoch 50.

In speech enhancement, T#9 (J+O→V+M) has a large JSUT–VCTK shift (Fig. 4b), yet GIRA still improves on zero-shot (MAE 10.84→10.61, SegSNR 13.22→13.46 dB). It also has the lowest MAE among source-free methods, and only the source-using CDAN and DANN reach a lower MAE.

In short, GIRA remains beneficial under moderate violations of the assumption and degrades only slightly under a strong one.

Related Revisions

  1. Secs. 3.3–3.4 (C1).
  2. Sec. 3.7 (C2).
  3. New App. C.6 (C3).
  4. Sec. 4.5 and Table 4 (C4).
Reviewer_CxptOpenReview ↗

Responses to Author

Thank you for the detailed response and additional experiments.

AuthorsOpenReview ↗

Responses to Reviewer Cxpt

Thank you for taking the time to consider our response and additional experiments. Your questions helped us explain the role of GIRA’s two gradient paths more clearly.

Reviewer_gm9LOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GIRA for adapting pretrained sequence-regression models using unlabeled target data without source-data access. GIRA defines neighborhoods in the model’s continuous prediction space as generalized classes and learns a conditional generator that maps these predictions to target-style inputs. It then adapts the regressor through the frozen generator, using uncertainty filtering, parameter regularization, and feature-scale constraints to stabilize training. Experiments cover nine transfer scenarios across non-intrusive load monitoring, electric load forecasting, and speech enhancement

Critical strengths

I find the use of a learned inverse model interesting because it provides an adaptation signal for continuous outputs, where common classification-based objectives are less suitable. This gives the paper a clear motivation for addressing source-free sequence regression.

Critical weaknesses
  1. My main concern is whether the adaptation objective corrects prediction errors or primarily enforces consistency with the original model. The generator learns from source-model predictions, and Equation (12) retains those predictions as fixed targets. This provides no independent evidence that the targets are correct. The source-quality experiments are helpful, but the paper needs a clearer explanation of when the resulting updates improve target accuracy rather than preserve systematic errors. This uncertainty affects the justification for the central adaptation mechanism.
  2. The benefit of learning a generative inverse is not clearly isolated from simpler consistency training. The framework combines generated inputs, uncertainty filtering, parameter regularization, and feature-scale constraints. Although alternative generators are evaluated, that comparison does not establish whether a learned inverse is necessary. A controlled comparison with direct prediction consistency under ordinary target-input perturbations, using the same filtering and regularization, would help determine whether the additional generative stage provides the claimed advantage.
Questions for authors
  1. Under what conditions does Equation (12) correct target prediction errors rather than reinforce the initial model’s errors? Please explain what property of the learned inverse makes the update direction useful, particularly when the initial predictions are confidently wrong.

  2. Do the existing experiments compare GIRA with direct prediction consistency under ordinary input perturbations, while keeping uncertainty filtering, parameter regularization, and feature-scale constraints identical? This would help isolate the contribution of learning a generative inverse.

  3. What happens if gradients through the conditioning prediction N(xT)N(x^T) in Equation (12) are stopped? Do any existing ablations establish whether this gradient path is essential to the reported gains?

  4. How were generator checkpoints, adaptation duration, and loss coefficients selected without labeled target data? Please distinguish the deployable selection procedure from analyses that use target labels retrospectively.

6 comments
AuthorsOpenReview ↗

Responses to Reviewer gm9L

We thank the reviewer for the careful review and for recognizing the value of a learned inverse for continuous outputs. We respond to the concerns below.

C1. When does Eq. (12) correct errors rather than reinforce the source model (W1, Q1)?

We thank the reviewer for raising this central point. To make the mechanism easier to follow, GIRA can be compared with self-supervised pseudo-labeling in classification SFDA (e.g., SHOT). In SHOT, each latent class contains many real target inputs. A centroid summarizes them, and every real input is reassigned to the centroid it resembles, so the correctly labeled majority of a class corrects its mislabeled minority. GIRA defines a generalized class (GC) in output space: the source prediction yˉ\bar{y} of a real target input serves as its identifier, and the GC contains the target inputs whose source predictions lie within τ\tau of yˉ\bar{y} (Eq. 4). We call the GC assignment of an input correct when its true output also lies within τ\tau of the identifier. Because each identifier is produced by one real input, a GC has no other real members that could outvote a wrong assignment.

This is what the "learned inverse" provides. For every identifier, GG generates new target-style inputs, and Eq. (12) pairs each of them with the identifier of its GC. UTF removes the most uncertain, and thus most likely wrong, pairs before GG is trained, so most generated pairs are correctly assigned. In Step 2, NN is fitted to the whole population, so a wrongly assigned generated input is predicted like the correctly assigned generated inputs it resembles, and its prediction moves toward its true output rather than toward the wrong identifier. Self-training on the real pairs cannot do this, since each GC then has a single member (see C2). Since the GCs are defined by the model itself, Eq. (12) also corrects their positions (see C3).

Eq. (12) therefore corrects errors when GG reproduces the target inputs accurately and most generated pairs are correctly assigned. GIRA improves on zero-shot once the source model is trained for 15 of 50 epochs (Table 17) and the generator for 3,000 steps (Table 5). A confidently wrong prediction passes UTF, but its generated pairs remain a minority among the correctly assigned generated inputs they resemble, so they are corrected rather than reinforced.

In addition, App. A.3 explains the mechanism of GIRA from a visual perspective.

C2. Does GIRA outperform direct prediction consistency under the same filtering and regularization (W2, Q2)?

We appreciate this suggestion and ran this controlled comparison (see Ru2P-C2). With UTF, LREG\mathcal L_{REG}, LSCC\mathcal L_{SCC}, the optimizer and the budget unchanged, replacing LGC\mathcal L_{GC} by Gaussian-perturbation consistency d(N(xT+σϵ),yˉ(xT))d(N(x_\mathcal{T}+\sigma\epsilon), \bar{y}(x_\mathcal{T})) or by self-training on the UTF-retained inputs stays close to zero-shot, whereas GIRA reduces MAE by 30.41% on T#1 and 47.96% on T#4.

C3. What happens if the gradient through N(xT)N(x_\mathcal{T}) in Eq. (12) is stopped?

We thank the reviewer for this question. We change only the condition of GG in Eq. (12) and keep all other settings as in Table 4.

Condition of GGT#1 MAET#4 MAE
Zero-shot36.57498.13
Fixed yˉ(xT)\bar{y}(x_\mathcal{T})29.69282.83
Stop-gradient sg[N(xT)]\mathrm{sg}[N(x_\mathcal{T})]28.73314.72
N(xT)N(x_\mathcal{T}), full gradient (GIRA)25.45259.22
  1. Both variants still improve over zero-shot, so most of the gain comes from correcting the generated population (see C1).
  2. Unlike classification, where the classes are given, the GCs are defined by the model itself, so the position of an identifier can also be wrong. Conditioning GG on N(xT)N(x_\mathcal{T}) lets this position move with NN while the target yˉ(xT)\bar{y}(x_\mathcal{T}) stays fixed. The gradient then takes two paths. The shallow path, through the outer NN, corrects the wrongly assigned generated inputs. The deep path, through GG back to the inner N(xT)N(x_\mathcal{T}), moves the class position to where the generated population is consistent with its identifier. Only GIRA keeps this path, which is why it performs best.

C4. Label-free selection of checkpoints, duration and coefficients.

Generator checkpoints are selected by the label-free Score (Eq. 75). Adaptation always runs for 40 epochs, fixed empirically. The loss coefficients, UTF ratio and learning rate (Table 6) were tuned once on T#4 with Optuna, using target-pool MAE as the objective, and then kept fixed for the other eight scenarios and all four backbones. Apart from this one-time search, no step uses target labels, and we correct App. B.2 accordingly. Tables 5 and 15–18 use target labels only as retrospective diagnostics. Table 18 shows low sensitivity: within 0.75–1.25× of the nominal values, T#4 MAE stays within 1.8% of 259.22.

Related Revisions

  1. Secs. 3.3–3.4 (C1, C3).
  2. Sec. 4.5 and Table 4 (C2).
  3. New App. C.6 (C3).
  4. App. B.2 (C4).

Thank you for the detailed response and the new experiments. The controlled comparison (C2) and the stop-gradient ablation (C3) are helpful, and I appreciate the correction to App. B.2.

On C1, the explanation relies on wrongly assigned generated pairs being a minority among similar inputs. This covers isolated errors but not systematic ones, where a region of the output space is biased in the same direction. Since the generator is trained on source predictions, its pairs would then be consistently wrong. Could you report, as a retrospective diagnostic, (i) the fraction of UTF-retained generated pairs that are correctly assigned, and (ii) the error change after adaptation for target inputs that were confidently wrong initially, or a bias/variance breakdown of the error before and after adaptation?

On C4, since hyperparameters were tuned on T#4 with target labels, please (i) report the average improvement over the eight other scenarios separately, (ii) clarify whether baselines received equivalent tuning, and (iii) state whether the 40-epoch duration was chosen using target labels.

On C3, fixed conditioning outperforms stop-gradient on T#4 but not on T#1. Are these single runs? Results over multiple seeds would help, as the current explanation does not account for this reversal. For C2, please note as a limitation that only Gaussian perturbations were tested.

Thank you for the response. The controlled comparison and the stop-gradient ablation answer Q2 and Q3 reasonably well, though the ablation shows that the gradient path helps rather than being essential, since fixed conditioning already recovers most of the gain.

My main concern remains open. The argument in C1 depends on most generated pairs being correctly assigned and on confidently wrong predictions being a minority, but neither is measured, and it does not cover systematic bias in the source model. Hyperparameters were also tuned with target labels on T#4, the scenario with the largest gain. Please clarify whether the tuning pool overlaps the T#4 evaluation data, and report the other eight scenarios separately, ideally over multiple seeds.

I will keep my current rating for now and am open to raising it if the requested diagnostics support the claimed mechanism.

Reviewer_Ru2POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Generative Inverse Regression Adaptation (GIRA), a source-free domain adaptation (SFDA) framework for sequence regression tasks. Unlike conventional SFDA methods primarily designed for classification, GIRA addresses the challenges arising from continuous output spaces, where discrete class assignments and classification-oriented adaptation objectives are not directly applicable.

The proposed framework introduces generalized classes (GCs), which define neighborhoods in the continuous prediction space of a pre-trained regression model. A conditional generative model is then trained to learn an inverse mapping from these generalized classes to target-style input sequences. The generated input–GC pairs are subsequently used to adapt the pre-trained regression model without access to source data or target annotations. To improve adaptation stability, the authors introduce uncertainty threshold filtering (UTF), weight regularization, and feature scale constraints. The paper also provides a theoretical target-risk bound to motivate the proposed framework.

Experiments are conducted on three sequence regression tasks—non-intrusive load monitoring, electric load forecasting, and speech enhancement—covering nine domain adaptation scenarios. The proposed method achieves substantial improvements over the pre-trained models and outperforms the compared SFDA methods in most evaluation settings.

Critical strengths
  • The paper addresses source-free domain adaptation for sequence regression, which has received considerably less attention than classification-based SFDA. The authors correctly identify the limitations of directly applying classification-oriented adaptation techniques to continuous regression outputs. Developing an adaptation framework that does not depend on discrete class assignments is a meaningful research direction.
  • The central idea of learning a conditional inverse mapping from regression outputs to target-style inputs is interesting. By introducing generalized classes, the proposed method provides a way to construct adaptation signals in continuous output spaces. The combination of conditional generation and regression model adaptation offers a potentially useful alternative to conventional pseudo-labeling approaches.
  • The experiments cover three distinct application domains and nine transfer scenarios, including low-frequency step signals, periodic signals, and high-frequency speech signals. This diversity provides useful evidence that the proposed framework is not limited to a single application. The results are generally strong, particularly for non-intrusive load monitoring and electric load forecasting.
  • The authors recognize that unreliable generated samples can cause negative transfer and introduce uncertainty threshold filtering, weight regularization, and feature scale constraints to address this issue. The ablation studies demonstrate that these components can substantially affect adaptation performance. In particular, the analysis of generator training quality provides useful insights into the relationship between generative modeling and downstream adaptation.
Critical weaknesses
  • The generalized-class identifiers are derived from predictions of the original source-trained model, which may already contain systematic errors on the target domain. The conditional generator is trained to produce target-style inputs consistent with these predictions, and the regression model is subsequently adapted using the generated input-GC pairs. However, it is not sufficiently clear why this procedure should move the regression model toward the true target regression function rather than simply reinforce the original model's predictions. In particular, the adaptation objective in Eq. (12) uses the fixed initial predictions as supervision while conditioning the generator on the current model predictions. Although this construction introduces a nontrivial adaptation signal, the mechanism by which it corrects systematic prediction errors remains insufficiently explained. The theoretical target-risk bound in Theorem 3.1 is potentially useful, but it contains terms involving generated–target discrepancy and generalized-class error whose practical behavior is difficult to assess without target labels. Consequently, the analysis does not fully establish when the proposed adaptation is expected to improve target-domain accuracy.
  • The paper compares GIRA with several UDA and SFDA methods, including SHOT, NRC, AaD, SSA, MAPU, and SFHarmony. However, several baselines were originally designed for classification and require modifications to accommodate regression outputs. While these comparisons are informative, they do not fully isolate the benefit of the proposed generative inverse mapping. In particular, comparisons with simpler alternatives, such as direct prediction consistency, uncertainty-filtered self-training, or target-data regularization using the same stability constraints, would help determine whether the improvements primarily arise from the proposed inverse mapping or from the accompanying regularization mechanisms.
  • GIRA requires training a conditional generative model before adapting the regression model. The main experiments employ Conditional MeanFlow, and the analysis in Table 5 shows that generator training quality strongly influences downstream performance. This introduces substantial additional computational requirements compared with methods that directly adapt the regression model using unlabeled target data. Although the paper provides implementation details and evaluates different generator configurations, a clearer comparison of end-to-end computational cost, training time, and memory requirements against representative baselines is needed to establish the practical advantages of the proposed framework.
Questions for authors

Please see the weaknesses section.

2 comments
AuthorsOpenReview ↗

Responses to Reviewer Ru2P

We thank the reviewer for the detailed review and the positive comments on our problem setting and evaluation. We respond to the concerns below.

C1. Why should Eq. (12) move the model toward the true target function, and how informative is Theorem 3.1?

We thank the reviewer for raising this central point. To make the mechanism easier to follow, GIRA can be compared with self-supervised pseudo-labeling in classification SFDA (e.g., SHOT). In SHOT, each latent class contains many real target inputs. The correctly labeled majority of a class corrects its mislabeled minority. In GIRA, the source prediction yˉ\bar{y} of a real target input serves as the identifier of a generalized class (GC), which contains the target inputs whose source predictions lie within τ\tau of yˉ\bar{y} (Eq. 4). The GC assignment of an input is correct when its true output also lies within τ\tau of the identifier. Because each identifier is produced by only one real input, GG supplies the missing class members by generating many target-style inputs for every identifier, and Eq. (12) pairs each of them with that identifier. Most generated pairs are correctly assigned, so when NN is fitted to the whole population, a wrongly assigned generated input is predicted like the correctly assigned generated inputs it resembles, and its prediction moves toward its true output rather than toward the wrong identifier. See gm9L-C1, gm9L-C3 and App. A.3 for details.

Theorem 3.1 identifies which quantities control the target risk. In its expanded form (Eq. 48), all data-dependent terms except the error–uncertainty envelope of Eq. (25) are label-free: LA\mathcal L_A, the generator Score (Eq. 75), and the UTF threshold and coverage γ\gamma. The bound is small when the generator matches the target inputs (small disc\mathrm{disc}) and the generated pairs are correctly assigned (small λh\lambda_h). Accordingly, GIRA beats zero-shot once the source model is trained for 15 of 50 epochs (Table 17) and the generator for 3,000 of 10,000 steps (Table 5). Conversely, it is unlikely to help when the generator is undertrained or the source model is too weak, since disc\mathrm{disc} or λh\lambda_h then dominates. With 1,000 or 2,000 generator steps, the T#4 MAE is 21.25% and 13.24% worse than zero-shot (Table 5), and with the epoch-10 source model 1.63% worse (Table 17).

C2. Do the gains come from the inverse mapping or from the accompanying regularization?

We appreciate this suggestion. To test this, we replaced LGC\mathcal L_{GC} by each alternative the reviewer lists, keeping UTF, LREG\mathcal L_{REG}, LSCC\mathcal L_{SCC}, the optimizer and the budget identical to Table 4. Each alternative uses its best weight among {0.25,0.5,1,2,4}×αGC\lbrace 0.25, 0.5, 1, 2, 4\rbrace\times\alpha_{GC} by target MAE, which favors the alternatives.

Adaptation termT#1 MAET#1 F1T#4 MAET#4 RMSE
Zero-shot36.570.52498.13584.75
None (UTF+LREG\mathcal L_{REG}+LSCC\mathcal L_{SCC})37.590.48498.77584.83
Perturbation consistency, σ=0.01\sigma=0.0136.850.52494.60582.53
UTF self-training toward yˉ\bar{y}36.730.49488.73579.05
LGC\mathcal L_{GC} (GIRA)25.450.67259.22334.96
  1. Without an adaptation term, the model stays near zero-shot, so the stabilizers alone do not produce the gains.
  2. Self-training toward the source identifiers, the objective that reinforces the original predictions, changes the results only slightly.
  3. Perturbation consistency recovers only a small part of GIRA's gain. This agrees with Table 1, where SHOT, NRC, AaD and MAPU, which all contain a Gaussian-view consistency term, are worse than zero-shot on T#1.

The gains therefore come from the inverse mapping rather than from the accompanying regularization.

C3. End-to-end computational cost, training time and memory.

We appreciate this suggestion. We now measure end-to-end time and peak GPU memory (maximum over T#1 and T#4) on the same RTX 4090. The time includes each method's preparation stage (e.g., Step 1 for GIRA). The full table with all baselines and the per-stage breakdown is given in the new App. C.7.

MethodT#1 time (min)T#4 time (min)Peak memory (GB)
SHOT21.3726.1112.9
SSA18.4222.2516.2
SFHarmony34.5545.1813.2
GIRA41.6445.3313.4
  1. Training the generator in Step 1 does make GIRA slower end to end, by about 23 min per task compared with SSA. In return, it lowers MAE by a further 27.8% on T#1 and 15.1% on T#4 relative to SSA, the strongest source-free baseline (Tables 1, 2).
  2. This extra time is spent only once before deployment. The generator is discarded after Step 2, so the deployed model keeps the source architecture and has the same inference cost as the other methods.
  3. GIRA's peak memory is comparable to the baselines.

Overall, this one-time cost is worthwhile given the clear accuracy gain.

Related Revisions

  1. Secs. 3.3–3.4 (C1).
  2. Sec. 3.7 (C1).
  3. Sec. 4.5 and Table 4 (C2).
  4. New App. C.7 (C3).
AuthorsOpenReview ↗

Responses to Reviewer Ru2P

Thank you again for taking the time to review our paper. As the discussion period draws to a close, we would appreciate hearing whether our earlier response and the additional experiments have addressed your concerns. If any questions remain or new concerns have come up, we would be happy to discuss them with you.

Comments on the paper

2 comments

Dear reviewers and authors,

With limited time left for discussion, please keep an eye on the newest replies and comments and deal with any outstanding concerns promptly. Thank you!

AC

AuthorsOpenReview ↗

Summary of the Review Discussion and Manuscript Revisions

Dear Program Chairs, Senior Area Chairs, Area Chairs, Reviewers

Thank you for overseeing the review. We also thank all three reviewers for their careful assessment and constructive suggestions. Their feedback helped us clarify how GIRA corrects prediction errors using unlabeled target inputs and explain what the experiments establish.

We appreciate the specific strengths highlighted in reviews.

  • Reviewer Ru2P recognized the importance of adapting sequence regressors without discrete class labels, the diversity of the nine transfer scenarios, and the strong results on appliance monitoring and load forecasting. We also appreciate the reviewer’s recognition of the analysis linking generator quality to downstream accuracy.
  • Reviewer gm9L highlighted the learned inverse as a way to obtain an adaptation signal for continuous outputs, where classification objectives are less suitable.
  • Reviewer Cxpt recognized the practical value of source-free sequence regression, the evaluation across three distinct tasks, and the ablations showing how individual components affect adaptation.

The central contribution of GIRA is to adapt sequence regressors without source data by defining generalized classes over continuous outputs and learning a conditional inverse mapping from these classes to target-style inputs. The pretrained regressor’s prediction on each retained target input serves as a class identifier. A generator trained on unlabeled target data produces input sequences for each identifier, and the resulting input–class pairs provide supervision for updating the regressor. We also derive a bound on the adapted regressor’s target risk. Across nine transfer scenarios, GIRA reduces mean absolute error (MAE) by 20.87% on average relative to the unadapted models and achieves the best task-average performance among the source-free methods compared on each of the three tasks (Tables 1–3).

The revised passages and new appendix text are marked in blue in the resubmitted PDF. In the main text, we clarified the two gradient paths in Eq. (12) (Secs. 3.3–3.4), explained how assignment error enters the target-risk bound (Sec. 3.7), and showed how the feature-scale constraint prevents a large output offset on T#4 (Sec. 4.5). These changes improve the presentation while retaining the original numerical results in Tables 1–3 and the paper’s core conclusions.

The appendix provides four main additions.

  • Appendix B.2 documents Optuna tuning by target MAE on the T#4 evaluation pool, label-free generator checkpoint selection, and the fixed 40-epoch budget. Appendix B.4 details the baseline grid searches on T#4 (Table 9).
  • Appendix C.3.4 uses an independent reference model for a retrospective diagnostic (Table 17). It estimates that 84.42% of generated pairs on T#1 and 94.26% on T#4 are correctly assigned after UTF filtering. After adaptation, 92% and 95% of initially confidently wrong real inputs had an estimated correct majority among their generated neighbors. MAE on the full subsets fell by 39.29% and 41.65%, respectively.
  • Appendix C.6 reports three-seed means showing that stop-gradient conditioning gives lower MAE than fixed conditioning on T#1 and higher MAE on T#4. Both improve on zero-shot, and the full gradient gives the lowest MAE on T#1, T#4, and T#5 (Table 20).
  • Appendix C.7 reports end-to-end training times on T#1 and T#4.

With the hyperparameters selected on T#4, GIRA reduced MAE by 17.49% on average across the other eight scenarios. Reviewer gm9L noted that the additional experiments addressed most of their concerns. At their request, the revision states that large coherent bias is unidentifiable without target-output information (Sec. 3.7), marks T#4 as the tuning scenario (Table 2), and gives the diagnostic tolerance in W² for T#1 and MW² for T#4 (Table 17).

Thank you for considering our responses and the resubmitted manuscript.

Sincerely, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.