acceptodds
Under review as a conference paper at ICLR 2027

EquiRFF: Radio Frequency Fingerprint Identification via Physical Equivariance Guided Representation Learning

Abstract

Radio frequency fingerprint identification (RFFI) seeks to preserve device-specific hardware cues under changing acquisition conditions, yet these cues are often entangled with domain factors in radio frequency (RF) signals. Overly strong invariance may suppress useful structural cues, while purely supervised models can exploit source-domain shortcuts. We propose EquiRFF, a two-stage framework that learns stable device identity from physical equivariance without explicit nuisance-factor separation. Stage I learns how structured physical transformations act in latent space and enforces consistency between sequential and composed transformations. Stage II extracts device-discriminative information that remains stable across these equivariant representations while retaining transformation information in the encoder. Physical views and condition parameters are used only during training. At inference, EquiRFF requires only a single I/Q observation, without physical-condition inputs or target-domain adaptation. Experiments on four public RF datasets show strong cross-domain performance against recent strong baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Predictive Representation Learning, Action-Conditioned Latent Prediction, Equivariant and Invariant Representation Learning, Radio Frequency Fingerprint

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_9R6tOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates cross-domain radio frequency fingerprinting using only source-domain data, aiming to address the challenge where device hardware fingerprints are entangled with variations in the communication channel, receiver, and acquisition environment. To tackle this, the authors propose a two-stage representation learning framework called EquiRFF. In the first stage, the framework learns how structured physical transformations operate within the latent space, enforcing consistency between sequential and composite transformations. In the second stage, it extracts device-discriminative information—which remains stable across transformations—from the previously learned equivariant representations, while simultaneously preserving transformation-related information within the encoder. Experiments conducted on four public datasets—UAV, ManySig, ORACLE, and BLE31—demonstrate that this method outperforms existing baselines.

Critical strengths

This manuscript proposes incorporating a physics-guided equivariant world model into radio-frequency fingerprinting. By employing an equivariant encoder alongside invariant cross-attention query heads, the approach enables one-shot inference without the need for further adaptation. The research addresses a significant problem, and the experimental results are promising.

Critical weaknesses

(1) Although extensive comparisons were conducted across four public datasets, the depth of the experiments should be increased. This could include, for example, visual verification of equivariance/invariance in the latent space, comparisons of computational overhead, validation using multiple random seeds, and a control experiment involving supervised training solely during Stage II.

(2) Is there a logical issue in treating a device’s inherent hardware fingerprint as an environmental disturbance that needs to be removed?

Questions for authors

(1) The experimental evaluation requires strengthening—for instance, by including visual verification of equivariance/invariance in the latent space, comparisons of computational overhead, validation across multiple random seeds, and a control experiment involving only the Stage II supervised training.

(2) How does the equivariance of noise-free, static CPO/CFO/IQI facilitate model generalization to distance-induced shifts—such as those dominated by multipath fading and path loss (as seen in the UAV and BLE31 datasets)?

(3) The baselines included in the manuscript consist primarily of general self-supervised visual learning methods. Furthermore, the manuscript lacks a standard supervised learning baseline—which should employ the same 6-layer Transformer architecture—and fails to clarify whether these self-supervised baselines also utilized the Stage II supervised contrastive loss (Lid=Lcls+LrffL_{id} = L_{cls} + L_{rff}) and the same physical data augmentation views during training.

(4) In the experiments on the BLE31 dataset (Table 4), most methods exhibit lower training accuracy than validation accuracy—a pattern that is inconsistent with typical training and testing outcomes.

3 comments

Response to Reviewer 9R6t — Part 1/1

We thank the reviewer for the detailed feedback

1 W1 + Q1 — Experimental depth

The revision adds physical diagnostics, supervised and from-scratch controls, cost measurements, and repeated-run statistics. Appendix Figures S1–S2 visualize physical endpoint prediction, encoder sensitivity, composition-path distance, and Encoder–Query geometry. Table S3 includes Supervised ERM (65.65 ± 0.05%), ERM + matched-scale augmentation (69.31 ± 0.05%), and Stage II from scratch (79.26 ± 0.41%). Tables S6–S8 quantify deployment and training costs, including 5.484M deployment parameters, approximately 0.677 GFLOPs, and 2.991 ms single-observation latency. The classification tables now report mean ± sample SD for seeds 2025, 2026, and 2027 across all four datasets and the UAV ablations. Appendix A.1 defines the aggregation, and Table S9 lists the seed-specific target averages.

Revision. Sections 4.3–4.4; Appendix A/B/D.

2 W2 — Hardware fingerprint preservation

EquiRFF is designed to preserve hardware information: it learns the response to an applied transformation, with the measured observation as its starting point. Relative recovery supervises transformation information in the encoder; classification and SupCon supervise the Query readout. This assigns physical response and stable identity to complementary representation roles. In the diagnostic checkpoints, all three factors show action-dependent endpoint prediction, and Query drift relative to device separation is lower than encoder drift. The w/o recovery and Encoder SupCon controls additionally test recovery and the location of identity consistency, yielding 86.55 ± 0.30% and 86.22 ± 0.21% compared with the Full reference of 95.51 ± 0.28%.

Revision. Sections 3.3.2 and 4.3; Appendix A/B/E.

3 Q2 — Distance, multipath, and path loss

Distance shifts are evaluated on real recordings, and we added synthetic multipath and noise tests: accuracy changes little under mild perturbations, while at the strongest settings (echo-to-direct energy ratio 1; received-signal-to-added-noise ratio 0 dB) it is 85.02% and 83.49%, compared with the clean reference of 95.77%. Both exceed the highest baseline accuracies of 67.44% and 69.54%; seq-JEPA has smaller relative drops from its lower clean reference. The training views are generated from real recordings, preserving their initial channel and receiver effects. Known phase, frequency, and I/Q transformations supply controlled supervision for the encoder, while identity learning extracts a consistent device representation across views, and the distance benchmarks evaluate this representation under the combined acquisition shift. To evaluate propagation filtering separately, the synthetic tests apply random short FIR channels and Gaussian noise to UAV target observations with fixed model weights; they quantify sensitivity to the specified channels, complementing the recorded distance shifts. Appendix F.5 reports the tested channel and noise strengths, while Table S14 compares all methods under the strongest FIR and added-noise settings. Appendix E.4 discusses operator selection and channel modeling.

Revision. Appendix E.4/F.5–F.6.

4 Q3 — Supervised baseline and downstream supervision

The advantage does not come from the downstream recipe alone, and a supervised baseline on the same Transformer is now included. In the matched Stage-II control, Vanilla 1D-JEPA with its CE classifier gives 86.78 ± 0.07%; adding the full EquiRFF Stage-II procedure gives 87.07 ± 0.43%, a 0.29-point increase in the mean. With the same Stage-II physical views, Query readout, classification, SupCon, and relative-recovery objectives, EquiRFF reaches 95.51 ± 0.28%, a further 8.44 points. This comparison supports the contribution of physics-based Stage-I pretraining under matched downstream training. Stage II from scratch gives 79.26 ± 0.41%; same-Transformer Supervised ERM and ERM + matched-scale augmentation give 65.65 ± 0.05% and 69.31 ± 0.05%, respectively. These controls further examine supervised training and physical augmentation without the proposed pretraining.

The main SSL baselines use physical views during pretraining and CE for downstream classification; Appendix C.2 lists their readouts and budgets. The matched Vanilla JEPA → Stage II control additionally includes EquiRFF's multi-view supervised procedure, SupCon, and relative recovery. AS retains its supervised and augmentation-adversarial contrastive objectives, and Matched-capacity MLP controls readout parameter count.

Revision. Sections 4.1 and 4.3; Appendix A/C/D.

5 Q4 — BLE31 Train and Val

Most methods on BLE31 show the usual ordering: eight of ten have a higher three-run Train mean than Val mean. The two exceptions are EquiMod and EquiRFF, whose Val means are higher by 0.39 and 0.22 points. For EquiRFF, Train and Val are 98.17 ± 0.83% and 98.39 ± 1.44%, and the paired Val−Train differences for seeds 2025–2027 are +2.71, −0.47, and −1.57 points: validation exceeds training in only one of the three runs, so the ordering varies across runs rather than showing a consistent validation advantage. Train and Val measure performance on different finite subsets, whose sample composition and fitted predictions can yield either ordering. BLE31 contains 31 devices; the 1 m source split uses 160 training and 40 validation observations per device, totaling 4,960 Train and 1,240 Val observations. Both mean accuracies are approximately 98%, near the accuracy ceiling, and one additional validation error changes accuracy by about 0.081 percentage points. The small mean gap should be read alongside these sample sizes and the observed run-to-run variation. Each held-out target distance contains 6,200 observations. Accuracies at 1.5, 2, and 3 m are 98.93 ± 0.41%, 96.94 ± 0.50%, and 98.14 ± 0.30%, with a target mean of 98.00 ± 0.08%. These target results are close to the source means and support strong identification across the tested distances. Appendix A.2 reports the three-run Train, Val, and target accuracy statistics.

Revision. Section 4.1 and Appendix A.2, including the BLE31 classification statistics.

Thank you for the author's response. Based on the content of the response and the full manuscript, I will maintain my score.

Reviewer_JegROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies radio frequency fingerprint identification across domains, where the task is to identify wireless devices when the receiver or the acquisition condition changes. The method has two stages. Stage I learns how physical transformations such as carrier phase offset, carrier frequency offset and I/Q imbalance change the latent representation, with an action encoding based on physics and a consistency loss between sequential transformations and the composed one. Stage II uses an attention query head to extract a device discriminative representation, and keeps transformation information in the encoder through a transformation recovery loss. At test time only one I/Q observation is needed, no adaptation on the target domain. The authors evaluate on four public datasets with 22 target domains and report the best accuracy on 19 of them. A cumulative ablation is also given.

Critical strengths
  1. The empirical results are strong: best accuracy on 19 of 22 target domains across four public datasets, covering receiver, distance, recording session and burst shifts. The gains on UAV reach more than 10 points on one domain, and they look consistent across the table, not from one or two lucky columns.

  2. The setting is practical: identification from a single I/Q observation, trained only on the source domain, without adaptation to the target domain. In real deployment there is usually no labeled data from the new environment, so this setting matters.

Critical weaknesses
  1. The two main technical ideas are close to earlier work. Homomorphism AutoEncoder [1] already learns latent dynamics where composition of transformations is respected, and AugSelf [2] already trains the model to predict the transformation between two augmented views. The paper does not cite or discuss either of them. Section 2.1 only distinguishes EquiRFF from V-JEPA 2-AC and seq-JEPA, but for me the closest works are [1] and [2], the first one for composition consistency and the second one for relative transformation recovery. Without this discussion, and ideally an empirical comparison, I cannot judge the added contribution.

[1] Keurti, Hamza, Hsiao-Ru Pan, Michel Besserve, Benjamin F. Grewe, and Bernhard Schölkopf. "Homomorphism Autoencoder--Learning Group Structured Representations from Observed Transitions." In International Conference on Machine Learning, pp. 16190-16215. PMLR, 2023.

[2] Lee, Hankook, Kibok Lee, Kimin Lee, Honglak Lee, and Jinwoo Shin. "Improving transferability of representations via augmentation-aware self-supervision." Advances in Neural Information Processing Systems 34 (2021): 17710-17722.

  1. It is hard to check the fairness of the comparison, and the ablation does not isolate the components. For the baselines, the paper only says they are adapted to one dimensional I/Q observations, and there is no description of the augmentations, the transformation inputs, the classification heads or the training budget of each method. The ablation in Table 5 is cumulative only, it never removes one component from the full model, so I cannot tell the individual effect of composition consistency or of the recovery loss, and in fact the recovery loss does not appear in the ablation at all.
Questions for authors
  1. What is the difference between your composition consistency and the transitions with group structure in [1], and between relative transformation recovery and the augmentation prediction in [2]? I mean the practical difference, not only the conceptual one. What do these differences change in the experiments?

  2. How were the vision baselines adapted to I/Q signals? Please describe the augmentations, the transformation inputs, the classification head and the training budget for each method, so the reader can check whether each adaptation keeps the key mechanism of the original method.

  3. Were the raw parameters normalized before the MLP in the raw action ablation? How many runs are behind each number? Please report the mean and the standard deviation over independent runs.

  4. Could you test the full model with one component removed at a time, especially composition consistency and the transformation recovery loss? Table 5 only adds components one by one, and the recovery loss is not ablated at all.

  5. Is there a control where Stage II is trained from scratch, without Stage I, but with the same physical views, the same Stage II losses and query head, and a comparable total budget? This would separate the contribution of Stage I pre-training from Stage II training on multiple views. After Stage II, is there any evidence that the encoder features still contain the transformation information while the RFF representation stays stable across views?

3 comments

Response to Reviewer JegR — Part 1/1

We thank the reviewer for the detailed feedback.

1 W1 + Q1 — HAE and AugSelf

Section 2.1 now cites HAE and AugSelf, while Appendix E.3 discusses their connections to EquiRFF. HAE relates to EquiRFF's composition design and AugSelf to its recovery objective; an AugSelf-style recovery control is also included in the ablations.

AugSelf and the recovery ablation. AugSelf predicts augmentation-parameter differences from two augmented views. Raw single-step recovery provides a closely related RF control within our framework: it predicts standardized raw parameters of a single relative RF transformation from adjacent encoder representations, retaining the same Stage-I initialization and Stage-II identity-learning setup as Structured single-step recovery. Structured single-step recovery instead predicts standardized structured physical coordinates, using the same adjacent view pairs. Their accuracies are 86.56 ± 0.26% and 91.10 ± 0.25%, respectively, a 4.54-point gain for structured recovery. Thus, this AugSelf-style parameter-prediction control tests the practical value of the recovery coordinates within EquiRFF. Appendix A.3 defines the selected view pairs and the separate preprocessing of conditioning inputs and recovery targets.

HAE and composition. The distinction from HAE is between learning a latent group action and using known RF operator composition to supervise latent prediction. HAE learns an action representation in an autoencoder whose group structure is expressed through a homomorphism, schematically ρ(g2∘g1)=ρ(g2)ρ(g1)\rho(g_2\circ g_1)=\rho(g_2)\rho(g_1), together with equivariant latent transitions. EquiRFF starts from known RF composition: same-factor CPO/CFO increments add, and IQI operators compose by ordered matrix multiplication. With z=f(x)z=f(x) and physical descriptor pgp_g, its nonlinear predictor is trained to satisfy

\qquad P(z,p_g)\approx\bar f(T_gx).$$ Here $\bar f$ is the EMA target encoder. These are learned prediction constraints anchored to the transformed observation; the predictor is not parameterized as an exact group representation. EquiRFF further reuses physical descriptors for Stage-II recovery while its Query readout learns device identity. Removing composition reduces target accuracy from 95.51 ± 0.28% to 89.01 ± 0.25%, and Appendix B evaluates actual endpoint errors alongside prediction-path agreement. **Revision.** Section 2.1 and Section 4.3; Appendix A/E. ### 2 W2 + Q2 — Baseline adaptation to IQ All baselines are adapted to IQ data under a two-level protocol: common data and evaluation conditions for complete methods, and matched downstream controls for contribution attribution. All baselines except AS use the shared Transformer backbone; AS retains MACNN-S and its learning framework. IQ is represented by two real-valued channels, with temporal patches for Transformer inputs. SSL pretraining uses RF physical views organized as paired observations, shared transformations, or action-linked sequences according to the method; SER also uses temporal block correspondences. AS uses its method-specific IQ augmentation pool. Appendix C summarizes these adaptations and tabulates the baseline readouts, encoder update policies, classification objectives, and epoch limits. Vanilla, CARE, and V-JEPA 2-AC use Query–MLP classifiers; EquiMod, STL, and BYOL use pooled-feature MLPs; seq-JEPA and SER train MLPs on frozen features. These SSL classification paths use CE, while AS jointly trains its linear classifier with its contrastive objectives. The matched control makes the contribution explicit: Vanilla 1D-JEPA gives 86.78 ± 0.07% with its CE classifier and 87.07 ± 0.43% with the complete EquiRFF Stage II (+0.29 points). EquiRFF reaches 95.51 ± 0.28% under the same Stage-II physical views, Query readout, classification, SupCon, and relative recovery (+8.44 points). Thus the gain remains when downstream training is matched and the pretraining method changes. Appendix D reports the separately specified computational benchmarks. **Revision.** Section 4.1; Appendix C/D.

3 Q3 — Raw parameters and repeated runs

Yes. Raw parameters were standardized before the MLP, and each number is now a mean ± sample SD over three independent runs (seeds 2025, 2026, 2027). In the Raw-action ablation, the four-dimensional vector (ϕCPO,ϵCFO,aIQI,dB,ψIQI)(\phi_{\mathrm{CPO}},\epsilon_{\mathrm{CFO}},a_{\mathrm{IQI,dB}},\psi_{\mathrm{IQI}}) is standardized coordinate-wise before entering the condition MLP: r~j=(rj−μj)/(σj+10−6)\tilde r_j=(r_j-\mu_j)/(\sigma_j+10^{-6}). We estimate the statistics once from 100,000 actions drawn from the configured training action distribution (statistics seed 3407) and keep them fixed as model buffers. They use no target-domain observations and are not recomputed per minibatch.

Across seeds 2025, 2026, and 2027, Raw action obtains 68.17 ± 0.33% and structured action obtains 80.75 ± 0.27%, a 12.58-point difference in mean target accuracy; Full obtains 95.51 ± 0.28%. We report mean ± sample SD (ddof=1). Appendix A provides per-domain statistics, each seed's target average, and the preprocessing details.

The standardization addresses the different numerical scales of radians, normalized frequency, and decibels, giving Raw a scale-controlled input. The fixed, invertible affine map preserves the parameter information and can be absorbed into the MLP's first linear layer. It changes the optimization parameterization while preserving the MLP's representable functions. Standardization operates on the action-conditioning path; signal transformations use the original physical parameters. The structured variant also uses coordinate-wise standardization and explicitly represents CPO/CFO through identity-centered circular coordinates and IQI through operator entries relative to the identity. The comparison therefore evaluates these encoding choices with coordinate scaling controlled.

Revision. Sections 4.1 and 4.3, Table 5; Appendix A.1/A.3/A.6.

4 W2 + Q4 — Removing composition and recovery

Removing either objective independently lowers the UAV target mean from 95.51 ± 0.28% (Full) to 89.01 ± 0.25% without composition (w/o composition removes the Stage-I composition objective) and to 86.55 ± 0.30% without recovery (w/o recovery removes Stage-II recovery). These independent removals complement the cumulative configurations in Table S2. Table S3 gives the complete domain results, and Table S4 directly checks the physical prediction behavior of the two-stage model.

As for the classifiers, configurations evaluated after Stage I alone use a fully connected (FC) classifier on the encoder representation. Configurations that include Stage II use the Query or MLP readout specified by the corresponding configuration, following the method description. The FC classifier used to evaluate Stage-I features is distinct from the MLP readout evaluated within Stage II.

Revision. Section 4.3; Appendix A/B.

5 Q5 — Stage II from scratch and preservation of physical information

Stage II from scratch, trained with the same total update budget as Full, gives 79.26 ± 0.41%, and Vanilla JEPA pretraining followed by Stage II gives 87.07 ± 0.43%, compared with 95.51 ± 0.28% for Full. These results examine the role of pretraining and its physical conditioning under the reported protocol.

For the reported UAV accuracy experiments, the Stage II from scratch variant trains from random initialization for 200 epochs, while Full uses 100 Stage-I epochs followed by 100 Stage-II epochs. Both use the same source training split, batch size 128, and gradient accumulation of one, matching the total optimizer-update budget. Stage II from scratch retains the Full model's Stage-II physical views, Query head, classification, supervised contrastive, and recovery objectives; source-validation performance selects the checkpoint.

After Stage II, the encoder still carries transformation information, and the RFF (Query) representation drifts less across views than the encoder. In the Stage-II diagnostic checkpoints listed in Table S4, correct actions are better than shuffled and identity conditions for CPO, CFO, and IQI. The target Query distance ratios are 0.043996, 0.021498, and 0.136711, all below the corresponding encoder ratios. These measurements characterize prediction and identity geometry in the evaluated checkpoints. IQI retains a 25.41% direct-error reduction over copying and a 4.97% recursive-error reduction. Under the tested IQI transformation, target classification accuracy changes from 95.77% to 95.68%, a decrease of 0.09 percentage points. Appendix B reports these complementary measurements together.

Revision. Sections 4.3–4.4; Appendix A/B.

Thank you for the authors' rebuttal response. While the responses have helped clarify several points, they have not substantially changed my overall assessment of the paper's significance. I therefore intend to maintain my original score.

Comments on the paper

6 comments
AuthorsOpenReview ↗

Update on Supplementary Experiments and Manuscript Revision

Dear Reviewers and Area Chairs,

We have uploaded an updated PDF that includes all additional experiments and analyses in the appendix, addressing the reviewers' comments and questions.

The main text has not yet been fully revised to reflect these additions. We are working on integrating the new results and clarifications into the main paper and aim to upload the revised version by November 16.

Thank you for your time and consideration.

We thank the reviewers for their questions on contribution attribution, physical diagnostics, and repeated-run evaluation. We report classification results for seeds 2025, 2026, and 2027 across four datasets and 15 UAV ablation configurations, with mean and sample standard deviation. EquiRFF obtains 95.51 ± 0.28%, 89.34 ± 0.20%, 95.54 ± 1.43%, and 98.00 ± 0.08% on UAV, ManySig, ORACLE, and BLE31, respectively. Component removals and training controls examine the roles of physical pretraining, composition, recovery, and the Query readout. Direct endpoint and identity diagnostics, together with training and deployment measurements, connect these design choices to representation behavior and cost.

DatasetEquiRFFStrongest baseline by target meanBaselineGap (pp)Highest domain mean
UAV95.51 ± 0.28Vanilla 1D-JEPA86.78 ± 0.078.7210/10
ManySig89.34 ± 0.20Avoiding Shortcuts84.43 ± 0.374.917/8
ORACLE95.54 ± 1.43Avoiding Shortcuts93.38 ± 0.172.161/1
BLE3198.00 ± 0.08Vanilla 1D-JEPA96.14 ± 0.151.862/3

Additional UAV stress tests use the seed-2025 EquiRFF model reported in the main comparison and nine seed-2025 baseline classifiers on the same 8,177 observations from ten target domains. We freeze all weights and reuse identical perturbation draws across methods. Classification uses the trained classifiers, and the additional tests report identification accuracy and changes from each clean reference. Target accuracy is averaged equally across domains, with a clean reference of 95.77% for EquiRFF. Appendix F reports these single-seed tests separately from the three-seed main comparison.

Response to Reviewer NqZn

We thank the reviewer for the detailed feedback.

1 W1 — Methodological contribution

EquiRFF differs from EquiMod, CARE, STL, and seq-JEPA in that it connects known RF transformation laws to identity learning through a shared physical descriptor. CPO and CFO use identity-centered circular coordinates; IQI uses the deviation of its I/Q operator from identity. Stage I conditions latent prediction on these coordinates and supervises same-factor compositions through their physical endpoint. Stage II recovers relative physical coordinates from encoder features while a separate Query readout learns device identity. This links physical prediction, composition, and identity learning through the same description of transformation.

Appendix E.2 compares all baselines using common mathematical notation. In particular, EquiMod models augmentation-conditioned embedding changes, CARE constrains representation geometry, STL infers transferable transformation representations, and seq-JEPA predicts from observation–action histories. EquiRFF connects known RF composition, endpoint prediction, and identity-stage recovery through the same physical descriptor.

The new controls support the physical conditioning and recovery design beyond a generic predictive encoder. Vanilla JEPA followed by the same Stage II gives 87.07 ± 0.43% mean target accuracy, compared with the Full result of 95.51 ± 0.28%. Independently removing composition or recovery gives 89.01 ± 0.25% and 86.55 ± 0.30%, respectively. Structured and raw single-step recovery give 91.10 ± 0.25% and 86.56 ± 0.26%. Appendix A reports every target domain; Appendix E gives the formulation.

Revision and supporting sections. Sections 2.1, 3.2.1–3.2.2, 3.3.2, and 4.3; Appendices A and E.1–E.2.

2 W2 — Physical prediction and unseen transformations

We now evaluate physical equivariance directly: the predictor's outputs are action-dependent and move toward the actual transformed endpoint, rather than merely satisfying composition consistency. For each checkpoint, we evaluate its observation encoder, action-conditioned predictor, and target encoder with all weights fixed. The diagnostic compares correct, shuffled, and identity actions against the encoding of the same actually transformed signal, and compares direct and recursive predictions with copying the initial encoder representation. For target observations after Stage II, direct prediction errors divided by copy errors are 0.357345, 0.492430, and 0.745907 for CPO, CFO, and IQI. Recursive ratios are 0.501228, 0.609317, and 0.950267. In the diagnostic checkpoints listed in Table S4, correct actions outperform shuffled and identity conditions for every factor in both stages and in both domain groups, which jointly demonstrates action dependence and useful prediction toward the physical endpoint. The effect of Stage-II recovery on downstream identification is evaluated separately through the w/o recovery and single-step recovery ablations in Tables 5 and S3.

Identification stays close to its three-step level under longer trajectories, and strictly out-of-range IQI changes accuracy little, whereas out-of-range CFO causes a clear drop. We evaluate four- to six-step trajectories and per-step actions outside the training bounds (CPO ±180°, normalized CFO ±5×10⁻⁴, IQI ±1 dB/±5°; three training steps). At six steps, CPO/CFO/IQI classification accuracies are 96.09%/94.17%/95.65%. Strictly out-of-range actions have CFO magnitudes in (5×10⁻⁴,10⁻³) or IQI magnitudes in (1,2) dB and (5,10)°, with random signs; three-step CFO/IQI accuracies are 83.05%/95.42%, corresponding to drops of 12.72/0.35 percentage points from the clean reference of 95.77%. These tests quantify identification performance under longer transformation sequences and unseen per-step magnitudes. Appendix F.2–F.3 specifies the trajectory lengths and action ranges.

Revision. Sections 4.3–4.4; Appendix B, Tables S4–S5, Figures S1–S2; Appendix F.1–F.3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.