acceptodds
Under review as a conference paper at ICLR 2027

FastAdapter: Codeword-Conditioned Adapter Generation for Cross-Encoder Alignment in CSI Feedback

Abstract

Independently trained CSI feedback encoders can produce incompatible codeword representations, degrading reconstruction by a fixed base-station decoder. Training an alignment model for each new encoder incurs repeated optimization. This paper proposes FastAdapter, which generates a complete adaptation module from unlabeled codewords received at the base station while preserving both endpoints. Offline, codeword refinement and reconstruction constraints prepare decoder-compatible target Adapters, and hidden-unit realignment gives functionally equivalent parameters consistent coordinates. A Gram–Raw condition encoder combines codeword coordinates, magnitudes, and inter-sample geometry to construct position-specific context. Contrastive pretraining associates this context with target parameters, and conditional diffusion generates reusable Adapters for unseen encoders. Evaluations on COST2100, WAIR-D, and DeepMIMO cover unseen training configurations and unseen architectures under a shared CSI distribution and feedback interface. Against the strongest generative baseline in each setting, FastAdapter lowers NMSE by – dB for unseen training configurations and – dB for unseen architectures. The generated Adapter reconstructs subsequent feedback without paired CSI supervision or gradient updates during adaptation, transferring offline adaptation experience across encoder models.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
CSI Feedback, Cross-Encoder Alignment, Adapter Generation, Diffusion Models

Reviews

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

This paper studies incompatibility between independently trained CSI feedback encoders and a fixed base-station decoder. It proposes FastAdapter, which generates a reusable intermediate adapter from unlabeled support codewords while keeping both endpoints frozen. Offline, the method prepares supervised adapter targets using codeword refinement and reconstruction constraints, then aligns their hidden-unit permutations to obtain consistent parameter coordinates. A condition encoder combines raw codewords, Gram-matrix features, and parameter-position information. Contrastive pretraining and conditional diffusion subsequently learn to generate complete adapter parameters for unseen encoders.

The evaluation covers COST2100 indoor/outdoor, WAIR-D, and DeepMIMO, using held-out training configurations and leave-one-architecture-out experiments over seven encoder families. FastAdapter reports lower NMSE than the strongest generative baseline in all eight settings, with improvements of 0.34–2.07 dB for unseen training configurations and 0.17–1.17 dB for unseen architectures. Component ablations, functional-equivalence checks for parameter realignment, and an adaptation-time breakdown provide additional validation.

Critical strengths

1.

This work tackles a meaningful interoperability problem with a practical adaptation interface.

Generating adapters from received codewords is a compelling approach, particularly when encoders and decoders are trained independently and must remain fixed. The framework cleanly separates the paired supervision used for offline preparation from the unlabeled observations available when adapting a new encoder, and the resulting adapter can be reused for subsequent feedback. Consistent improvements across both training-configuration and architecture holdouts offer promising evidence that adaptation experience transfers across different encoder models. This is the core reason I see the problem formulation and proposed system as genuinely valuable.

2.

Key design choices are supported by targeted experiments and a dedicated correctness check.

The paper addresses parameter-coordinate inconsistency as a prerequisite for learning to generate adapter weights. Appendix A demonstrates that jointly permuting the two FFN weight matrices and the first-layer bias preserves the adapter function, and Table 4 confirms this result numerically. Table 2 further shows a substantial drop in reconstruction quality when ParamRealign is removed. The Gram and Raw branches, position-specific context, and contrastive pretraining also consistently boost performance across all evaluated settings. Taken together, these experiments make a solid case that both the target representation and conditioning design are meaningful contributors to the system’s performance.

Critical weaknesses
  1. The encoder population and the reliability of the reported generalization gains are insufficiently characterized.

Section 5.1 and Appendix F describe grouped splits and seven architecture families, but do not specify the actual training configurations varied or provide complete encoder counts for the main evaluations. These details are central because the method learns a mapping from codeword statistics to adapter parameters, so its generalization depends on the diversity and independence of the encoder tasks.

Furthermore, Table 1 reports aggregate point estimates without variability across support-set draws, diffusion initializations, or repeated generator training. Results for the individual held-out architecture families are also absent. This matters particularly for the smaller improvements of 0.17 dB and 0.34 dB. The current evidence does not establish whether these gains are reproducible or whether aggregate scores conceal substantial variation across encoder families. These omissions limit my confidence in the paper’s main generalization claim. References: Section 5.1 Table 1 Appendix F Section 6 also acknowledges unmeasured support-set effects and generation variability

  1. The incremental value of diffusion and its computational tradeoff are not adequately isolated

HyperTransformer is included, and the paper states that the baselines share target adapters and conditioning information. However, it does not document a matched deterministic predictor that retains the full Gram–Raw condition encoder, position-specific context, and target preparation while replacing only the diffusion generator. Consequently, the experiments do not clearly establish how much of the improvement requires diffusion rather than the proposed conditioning pipeline.

This comparison is consequential because 50-step DDIM sampling accounts for 4.48 seconds of the reported 5.99-second adaptation pipeline. Table 3 reports timings for adapter optimization, affine adaptation, and retrieval, but omits all four competing parameter generators. The paper therefore establishes an NMSE advantage over those generators without establishing the corresponding accuracy–latency tradeoff. A matched direct predictor and comparable generator timings would materially strengthen the justification for the additional generation cost.

my references: Section 5.1, Table 3, Appendix D, Table 8, Section 6.

Questions for authors
  1. Clarification of the encoder-generalization protocol. What exactly varies across the “training configurations,” and how many independently trained encoder instances are included in each dataset, architecture family, and split? How are checkpoints or models derived from the same training run grouped? Please also provide the concrete codeword dimensions and compression ratios used in the main experiments.

  2. Clarification of hyperparameter selection. Table 5 reports selected NMSE values of −21.49 and −21.18 dB, which exactly match the COST2100/in FastAdapter entries in Table 1. Which models and data splits produced the Appendix B scores: fitted target adapters or generated adapters, and validation or test encoders? Please clarify how the loss weights were selected and how the selection data were separated from the final evaluation.

  3. Validation of reliable adaptation from unlabeled support. Does Table 1 evaluate a single generated adapter per encoder without CSI-based selection among multiple candidates? How do the results vary across support-set draws and diffusion initializations, particularly for the smallest reported gains? Since support codewords are drawn from each encoder’s training CSI split, do the results persist when support CSI is independently drawn and was not used to train that encoder?

  4. Validation of the contribution of diffusion. With the target adapters, full condition encoder, position-specific context, and training budget held comparable, how does a deterministic parameter predictor perform in NMSE and adaptation time? What are the corresponding timings of HyperTransformer and the diffusion-based baselines on the same hardware?

3 comments
AuthorsOpenReview ↗

Response to Reviewer eyNq W1-1/Q1 & W1-2

Thank you for your thoughtful review and valuable feedback. We appreciate your recognition of the practical value of encoder interoperability and our reusable Adapter interface. Below are our responses to your comments (your comments are highlighted in italics).

W1-1: The encoder population and the reliability of the reported generalization gains are insufficiently characterized. [...] These details are central because the method learns a mapping from codeword statistics to adapter parameters, so its generalization depends on the diversity and independence of the encoder tasks.
Q1: Clarification of the encoder-generalization protocol. What exactly varies across the “training configurations,” and how many independently trained encoder instances are included in each dataset, architecture family, and split? How are checkpoints or models derived from the same training run grouped? Please also provide the concrete codeword dimensions and compression ratios used in the main experiments.

Thank you for asking how we count independent encoders. We select one checkpoint per run by validation performance. Its target Adapter and codewords share its split.

In preliminary tests within each family and dataset, the training seed had the largest effect. We vary it, fixing batch size, peak learning rate, schedule, and optimizer. Each dataset has 1213 encoders per family (8491 total). We hold out instances within families or one family per fold (Table 1). No Adapter yields 23.07–29.32 dB on unseen configurations.

Table 1. Encoders per dataset.

DatasetUnseen training configurations: train/val/test per familyUnseen training configurations: train/val/test across seven familiesUnseen architectures: test encoders per fold
COST2100/in973/121/1196811/847/8331213
COST2100/out973/121/1196811/847/8331213
WAIR-D973/121/1196811/847/8331213
DeepMIMO973/121/1196811/847/8331213

Table 2. CSI and codeword dimensions and compression ratios.

SettingCOST2100/inCOST2100/outWAIR-DDeepMIMO
Real-valued CSI input dimensions2×32×322×32×322×64×642×32×64
Codeword dimension C51251220481024
Compression ratio CR1/41/41/41/4

In CR=C/(2HW)CR=C/(2HW), the two channels are real and imaginary. All families in a dataset share the codeword dimension and decoder interface.

W1-2: Furthermore, Table 1 reports aggregate point estimates without variability across support-set draws, diffusion initializations, or repeated generator training. [...] Results for the individual held-out architecture families are also absent. [...] This matters particularly for the smaller improvements of 0.17 dB and 0.34 dB.

Thank you for suggesting repeated runs and per-family results. Across three new runs, FastAdapter has the lowest mean NMSE among the compared generators in all eight settings.

  1. The repeats cover training and generation. Seeds 42, 2026, and 3407 vary generator initialization and training, support and query draws, and DDIM initial noise. We fix the split (seed 0), encoders, target Adapters, alignment reference, and standardization statistics. Each repeat generates one Adapter per test encoder. For each encoder, we sum error and channel energies over queries and convert their ratio to dB. We average within families, then equally across families, and report mean±std over repeats.

  2. NMSE varies by held-out family. TransNet and Attention-CNN reach -16.60 and -8.99 dB on COST2100/out (Table 3).

  3. The small gains persist on average. Gains over RPG are 0.17 dB on COST2100/out (unseen architectures) and 0.27 dB on WAIR-D (unseen configurations). On DeepMIMO, the gains over D2NWG are 0.67 dB (unseen configurations) and 0.59 dB (unseen architectures), with a 0.22 dB lead over Flow Matching in the latter case (Table 4).

Table 3. FastAdapter-50 NMSE by held-out family (dB).

Encoder familyCOST2100/inCOST2100/outWAIR-DDeepMIMO
CsiNet-21.34±0.15-13.15±0.18-20.37±0.07-20.12±0.02
CRNet-22.26±0.20-14.70±0.06-21.10±0.16-22.13±0.12
CLNet-23.47±0.05-13.98±0.09-22.30±0.18-21.39±0.07
TransNet-24.74±0.11-16.60±0.07-23.28±0.14-23.43±0.04
ResNet-19.88±0.21-10.40±0.04-17.84±0.19-18.66±0.02
MLP-AE-19.06±0.22-11.44±0.08-18.70±0.08-17.61±0.07
Attention-CNN-17.36±0.12-8.99±0.20-16.34±0.23-16.18±0.14
Avg.-21.16±0.14-12.75±0.07-19.99±0.10-19.93±0.05

Table 4. NMSE and mean gains (dB).

ProtocolDatasetComparison methodBaseline NMSEFastAdapter-50Mean improvement
Unseen training configurationsWAIR-DRPG-20.78±0.05-21.05±0.150.27
Unseen training configurationsDeepMIMOD2NWG-20.43±0.16-21.10±0.130.67
Unseen architecturesCOST2100/outRPG-12.58±0.11-12.75±0.070.17
Unseen architecturesDeepMIMOFlow Matching-19.71±0.15-19.93±0.050.22

In the uploaded revised manuscript, see Appendix G.1 for the protocol and G.2 for three-run and per-family NMSE.

AuthorsOpenReview ↗

Response to Reviewer eyNq W2 & Q4

W2: The incremental value of diffusion and its computational tradeoff are not adequately isolated. [...] However, it does not document a matched deterministic predictor that retains the full Gram–Raw condition encoder, position-specific context, and target preparation while replacing only the diffusion generator. [...] Table 3 reports timings for adapter optimization, affine adaptation, and retrieval, but omits all four competing parameter generators.
Q4: Validation of the contribution of diffusion. With the target adapters, full condition encoder, position-specific context, and training budget held comparable, how does a deterministic parameter predictor perform in NMSE and adaptation time? What are the corresponding timings of HyperTransformer and the diffusion-based baselines on the same hardware?

Thank you for suggesting that we compare both generator accuracy and runtime. Under unseen training configurations, FastAdapter's mean NMSE is 7.43–8.35 dB lower than MLP-L's, but adaptation takes longer. Compared with Cond P-Diff, D2NWG, and RPG, FastAdapter has both accuracy and runtime advantages.

  1. Controls that replace only the generator directly test the contribution of the generation mechanism. MLP-S / MLP-L retain the full conditioning pipeline and replace only the generator, using the same training budget as FastAdapter. The HyperTransformer reimplementation is also deterministic: it predicts the Adapter parameters used in this work and produces the same output for the same model parameters and support codewords. All compared methods use comparable hyperparameter-tuning budgets, with configurations selected by validation performance.
  2. FastAdapter reconstructs CSI more accurately than MLP-L, with a longer adaptation time. On COST2100/in under unseen training configurations, FastAdapter takes 2.92 s longer than MLP-L and reduces mean NMSE by 8.35 dB. Tables 5 and 6 report the accuracy and runtime of each generative method.

Table 5. NMSE of parameter-generation methods under unseen training configurations (dB).

MethodCOST2100/inCOST2100/outWAIR-DDeepMIMO
MLP-S-2.16±0.15-1.18±0.27-1.00±0.22-0.43±0.31
MLP-L-13.16±0.16-7.34±0.14-12.83±0.15-13.67±0.23
HyperTransformer-16.15±0.07-9.71±0.15-17.04±0.13-17.51±0.14
Cond P-Diff-19.89±0.08-12.46±0.11-19.22±0.18-17.13±0.09
D2NWG-20.67±0.08-12.83±0.06-18.39±0.12-20.43±0.16
RPG-20.25±0.07-13.49±0.12-20.78±0.05-20.40±0.11
Autoregressive-13.23±0.15-7.09±0.13-12.27±0.20-12.86±0.19
Flow Matching-20.31±0.22-13.47±0.19-20.43±0.17-20.36±0.13
FastAdapter-50-21.51±0.09-15.66±0.17-21.05±0.15-21.10±0.13

Table 6. Total adaptation time of generators under unseen training configurations (s).

MethodCOST2100/inCOST2100/outWAIR-DDeepMIMO
MLP-S1.71±0.121.72±0.131.82±0.121.76±0.15
MLP-L2.83±0.072.86±0.133.14±0.203.08±0.15
HyperTransformer2.91±0.812.86±0.773.51±0.923.14±0.89
Cond P-Diff45.98±5.4142.45±4.3650.09±4.9246.79±4.54
D2NWG92.98±7.4489.04±8.0388.76±7.4691.09±7.45
RPG103.92±7.8999.53±8.41120.72±10.49115.64±9.38
Autoregressive15.63±1.4714.28±1.8318.74±2.2917.52±1.94
Flow Matching34.72±3.9132.14±3.4739.83±4.2837.56±4.01
FastAdapter-505.75±0.186.08±0.156.16±0.186.06±0.29

Under unseen training configurations, FastAdapter takes 5.75–6.16 s, compared with 2.86–3.51 s for HyperTransformer and 42.45–120.72 s for Cond P-Diff, D2NWG, and RPG. On COST2100/in under unseen training configurations, DDIM takes 4.33±0.16 s out of a total adaptation time of 5.75±0.18 s, making sampling the main cost. Generation is performed when adapting to a new encoder, and subsequent feedback reuses the Adapter. All methods are timed on the same RTX 4090, from support-codeword encoding to the completion of Adapter loading. In the uploaded revised manuscript, Appendix G.3 describes the new deterministic controls, and Appendices G.2 and G.5 report accuracy and the new generator timings, respectively, for both protocols.

AuthorsOpenReview ↗

Response to Reviewer eyNq Q2 & Q3

Q2: Clarification of hyperparameter selection. Appendix B/Table 5 reports selected NMSE values of −21.49 and −21.18 dB, which exactly match the COST2100/in FastAdapter entries in Table 1. Which models and data splits produced the Appendix B scores: fitted target adapters or generated adapters, and validation or test encoders? Please clarify how the loss weights were selected and how the selection data were separated from the final evaluation.

We apologize for the confusion caused by our description of hyperparameter selection and the displayed results. The weights are selected using validation encoders; the 55-cell grid evaluates the NMSE of FastAdapter-generated Adapters on test encoders. In the uploaded revised manuscript, Appendix B and Section 5.2 report the same test evaluation for the selected weights, hence the identical NMSE.

For each candidate loss-weight configuration, we first prepare the corresponding target Adapters and train the generator. We then generate Adapters from the support codewords of validation encoders and select the weights by reconstruction NMSE. After selection, we fix the configuration and evaluate it on independent test encoders. The evaluation measures generated Adapters, not fitted target Adapters.

Table 7. Validation selection and test evaluation of generated Adapters, NMSE (dB).

ProtocolBest validation-grid NMSETest mean±std
Unseen training configurations-21.82-21.51±0.09
Unseen architectures-21.32-21.16±0.14

Appendix G.7 of the uploaded revised manuscript gives the expanded validation grid of 77 configurations per protocol. Among all configurations searched for each protocol, the lowest validation NMSE for both protocols is attained at (λalign,λrecon,λref)=(1,1000,2)(\lambda_{\mathrm{align}},\lambda_{\mathrm{recon}},\lambda_{\mathrm{ref}})=(1,1000,2). WAIR-D and DeepMIMO reuse these weights without a separate search over target-loss weights. After all target Adapters have been trained, we use seed 0 to fix the split of encoder tasks into training, validation, and test sets. All parameter-generation methods use the same split and are supervised by the target Adapter parameters corresponding to the training-set encoders.

Q3: Validation of reliable adaptation from unlabeled support. Does Table 1 evaluate a single generated adapter per encoder without CSI-based selection among multiple candidates? How do the results vary across support-set draws and diffusion initializations, particularly for the smallest reported gains? Since support codewords are drawn from each encoder’s training CSI split, do the results persist when support CSI is independently drawn and was not used to train that encoder?

Thank you for your attention to the reliability of adaptation from unlabeled support. Yes, we generate only one Adapter per encoder in each evaluation, without using query CSI to select among multiple candidates. When we instead use independent support CSI that was used neither for encoder training nor for checkpoint selection, the absolute change in mean NMSE is no greater than 0.09 dB across the four datasets under unseen training configurations.

  1. We generate only one Adapter per encoder in each evaluation and repeat the evaluation to assess the effects of support sampling and initial noise. For each test encoder, we draw 128 support codewords, sample the initial noise once, and use DDIM to generate one Adapter; query CSI is used only to compute the final NMSE. Each of the three repeats follows this procedure, jointly varying support sampling, initial denoising noise, and generator training. Under unseen training configurations, FastAdapter's standard deviations across the three repeats are 0.09–0.17 dB. The aggregation procedure and repeated results for the small margins are provided in our response to W1-2 and in Table 4.
  2. Independent support still achieves accuracy close to that of the main experiments. The independent support pool is used neither for encoder training nor for checkpoint selection and does not overlap with the query set. Under unseen training configurations, mean NMSE changes by +0.02 to +0.09 dB, and the standard deviations with independent support are 0.10–0.21 dB. In the in-distribution evaluation on these datasets, support sets formed from CSI outside encoder training and checkpoint selection still enable effective adaptation.

Table 8. NMSE with different support CSI sources under unseen training configurations (dB).

DatasetTraining-CSI supportIndependent-CSI support
COST2100/in-21.51±0.09-21.46±0.10
COST2100/out-15.66±0.17-15.64±0.18
WAIR-D-21.05±0.15-20.98±0.17
DeepMIMO-21.10±0.13-21.01±0.21

In the uploaded revised manuscript, Appendix G.1 describes the single-Adapter evaluation and repeat protocol, Appendix G.6 reports the new independent-support experiment under both protocols, and Section 6 discusses the implications for deployment.

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

The paper addresses codeword drift. In CSI feedback, when a user-equipment encoder is independently trained or updated, its codeword distribution no longer matches what the base station's fixed decoder learned, and direct connection degrades NMSE to 17.6–31.0 dB. FastAdapter proposes to generate the entire base-station-side Adapter from only the 128 unlabeled codewords the BS has already received. No paired CSI, no encoder weights, no architecture label.

First, offline target preparation: instead of regressing pointwise to a reference encoder's codewords, a candidate codeword is refined through the frozen decoder so that it reproduces the source pair's reconstruction, and a per-encoder Adapter is then fitted to that target. Defining the target as "what the decoder must receive" rather than "what the encoder emitted" is the point of this part. Second, ParamRealign: independently fitted Adapters can express the same function under different hidden-unit orderings, so Hungarian matching against a shared reference places the parameters in consistent coordinates. Without this alignment the generation targets are not well defined. Third, a Gram–Raw condition encoder summarizes codeword coordinates, magnitudes and inter-sample geometry into a per-parameter-token context. Fourth, contrastive pretraining ties condition to target parameters, after which conditional diffusion generates the Adapter tokens. At deployment, 50 DDIM steps produce an Adapter that is inserted between the unseen encoder and the fixed decoder, and subsequent codewords reuse it.

Evaluation covers seven encoder families, COST2100/WAIR-D/DeepMIMO, and two protocols (unseen training configuration, leave-one-architecture-out), for eight settings total. FastAdapter has the lowest NMSE among generative methods in all eight (0.17–2.07 dB over the strongest), trails a per-encoder supervised oracle by 0.96–2.65 dB, and adapts in 5.84–6.21 s against the oracle's 16,959–24,841 s. Ablations raise NMSE by 7.76–11.17 dB without ParamRealign and 9.62–13.66 dB without the global condition.

Critical strengths

The paper sets up its problem precisely and never leaves the information constraint it states. The base station holds only the codewords an unidentified encoder has uploaded. No paired CSI, no encoder weights, no architecture label. Fixing that constraint up front is what makes the question well posed: "why not use labels" never arises, and it is clear what has to be measured. Naming the deployment setting this explicitly is what makes the problem answerable at all, and the paper holds to it throughout.

What is technically distinctive is that two design choices are built as premises for each other rather than as independent tricks. The target is not defined as the codeword the encoder emitted but as the codeword that, passed through the frozen decoder, reproduces the source pair's reconstruction. The correction is defined in the decoder's input space rather than its parameter space, which is the only form that can be deployed over a single fixed decoder. On top of that, independently fitted Adapters are aligned to a shared reference so that the same function cannot be expressed in different coordinates. Without that alignment the generation targets are not well defined, and the ablation shows it: 7.76–11.17 dB.

The comparison design is fair. Four generative baselines are reimplemented in a shared token space with the same splits, the same target Adapters, the same codeword conditions and the same parameter manifest, alongside simple controls (direct connection, mean, nearest, affine) and a supervised oracle. Seven encoder families under two protocols give eight settings, and the paper reports the lowest NMSE among generative methods in all eight. The reporting stance is worth noting too: the runtime breakdown discloses that DDIM sampling accounts for 74.8% of adaptation time, the limitations section names by hand the two experiments that could refute the paper, and the absence of quantized feedback and of channel and hardware shifts from the evaluation is admitted outright. Application papers rarely state their own weaknesses this concretely.

Critical weaknesses
  1. The headline margins are single point estimates from a pipeline with three sources of randomness, and no variance is reported anywhere.

The pipeline draws randomness from three places: the DDIM initial noise, the random K = 128 support-codeword draw, and the encoder split. There are no seeds, no repeats, no error bars and no significance tests, and the limitations section itself concedes that "generation variability remain[s] to be measured." The margins are small enough that this matters. Computed from the paper's own table, the advantage over the strongest generative baseline in each of the eight settings is 0.80, 2.07, 0.34, 0.64, 1.17, 0.17, 1.13 and 0.53 dB — five below 1 dB and two below 0.5 dB. The encoder population behind each macro-average is not stated either, so differences of this size cannot be interpreted. The unhedged superlative "lowest NMSE among generative methods in all 8 settings" nevertheless appears in the abstract, the results text and the conclusion.

  1. The contribution of the core mechanism is never isolated, and diffusion accounts for three quarters of the deployment cost.

The ablation table covers condition construction and ParamRealign only. The one deterministic data point, HyperTransformer, is fully confounded: whether the reimplementation was deterministic is not stated, it was designed for few-shot CNN weights and so sits in a different domain from Adapter tokens, it is the worst baseline in all eight settings (3.55–6.96 dB behind FastAdapter's own result), and baseline tuning parity is unreported. Determinism and domain mismatch are entangled, so no conclusion about diffusion can be drawn from that row. The clean in-model control — the paper's own Gram–Raw condition encoder with a deterministic regression head under the same tokens, the same targets and the same training budget — appears nowhere, even though the limitations section names this experiment: "A deterministic generator trained with the same conditions would further clarify the contribution of diffusion." This is load-bearing because diffusion is not a minor component: DDIM sampling is 4.48 s of the 5.99 s pipeline, 74.8% of the total. If a deterministic head reached the same NMSE in roughly a quarter of the time, the contribution would have to be reframed around target preparation and condition encoding rather than around parameter diffusion.

Target preparation sits in the same position. The choice the paper argues for in prose — refining a candidate codeword through the frozen decoder rather than regressing pointwise to a reference encoder's codewords — is not ablated, and the target Adapter's own NMSE is reported nowhere, so the 0.96–2.65 dB gap to the supervised oracle cannot be attributed between preparation and generation.

Smaller things, not driving my recommendation: "training configuration", the split unit of the headline protocol, is never defined; Table 4 counts 973 training Adapters where the Figure 4 caption says 946; apart from the parameter-token width of 512, constants such as codeword dimension, compression ratio, Adapter depth, guidance strength, temperature, batch size and context width appear only as symbols with no values; "Online Affine" and "Affine Adapter" are the same method under two names; and all evaluation uses unquantized codewords drawn from the encoder's training split, while the appendix describes deployment as forming the support set from "previously received feedback codewords".

Questions for authors
  1. Variance. For the settings whose margins are under 0.7 dB, can you report NMSE mean ± std over multiple DDIM noise draws and multiple support-set draws, and over training seeds? Do the 0.17 dB margin over RPG on COST2100/out under the unseen-architecture protocol and the 0.34 dB margin on WAIR-D under unseen training configurations survive resampling?
  2. Deterministic control. What NMSE does your own Gram–Raw condition encoder reach with a deterministic regression head under the same tokens, the same targets and the same training budget, and what is its deployment time against the 5.99 s pipeline? The limitations section names this experiment directly — why is it not in the ablation table? Related: was the HyperTransformer reimplementation deterministic?
  3. Loss-weight search protocol. On which split were the 55 grid cells computed? The grid minimum matches Table 1's FastAdapter test NMSE to two decimal places under both protocols, and the appendix describes the grid metric as "NMSE for both evaluation protocols", while Section 5.1 states that validation selected the weights. Related: why is λ_align fixed at 1 and never searched, why does λ_ref = 2 sit at the top of the search range, and were the WAIR-D and DeepMIMO weights transferred from COST2100/in without validation? Since the selected weights define the target Adapter that every baseline consumes, the answer bears on the whole comparison rather than on one table.
4 comments
AuthorsOpenReview ↗

Response to Reviewer joww W1 & Q1

Thank you for your detailed review and helpful suggestions. We appreciate your recognition of how compatible targets and parameter realignment work together, as well as the fairness of our comparisons. Below are our responses to your comments (your comments are highlighted in italics).

W1-1: The headline margins are single point estimates from a pipeline with three sources of randomness, and no variance is reported anywhere. [...] The margins are small enough that this matters.

Q1: For the settings whose margins are under 0.7 dB, can you report NMSE mean ± std over multiple DDIM noise draws and multiple support-set draws, and over training seeds? Do the 0.17 dB margin over RPG on COST2100/out under the unseen-architecture protocol and the 0.34 dB margin on WAIR-D under unseen training configurations survive resampling?

Thank you for suggesting repeated evaluations. In the three runs added for this revision, FastAdapter achieves the lowest mean NMSE among the compared generative methods in all eight settings. Table 1 reports mean±std and the mean gain over the strongest generative baseline.

  1. The repeats cover both training and generation. Seeds 42, 2026, and 3407 vary generator initialization and training, support and query draws, and DDIM initial noise. The encoders, target Adapters, ParamRealign reference, standardization statistics, and task split (seed 0) are fixed. Each repeat generates one Adapter per test encoder from 128 support codewords. We report mean±std over the three family-balanced macro-averages.
  2. The four small margins are 0.17–0.67 dB. The mean improvements over RPG are 0.17 dB on COST2100/out under unseen architectures and 0.27 dB on WAIR-D under unseen training configurations. On DeepMIMO, the mean improvements over D2NWG are 0.67 dB under unseen training configurations and 0.59 dB under unseen architectures; under unseen architectures, the mean improvement over the strongest baseline, Flow Matching, is 0.22 dB.

Table 1. NMSE from repeated evaluation (dB) and mean improvement over the strongest generative baseline.

ProtocolDatasetStrongest generative baselineBaseline NMSEFastAdapter-50Mean improvement
Unseen training configurationsCOST2100/inD2NWG-20.67±0.08-21.51±0.090.84
Unseen training configurationsCOST2100/outRPG-13.49±0.12-15.66±0.172.17
Unseen training configurationsWAIR-DRPG-20.78±0.05-21.05±0.150.27
Unseen training configurationsDeepMIMOD2NWG-20.43±0.16-21.10±0.130.67
Unseen architecturesCOST2100/inFlow Matching-20.11±0.20-21.16±0.141.05
Unseen architecturesCOST2100/outRPG-12.58±0.11-12.75±0.070.17
Unseen architecturesWAIR-DD2NWG-18.86±0.13-19.99±0.101.13
Unseen architecturesDeepMIMOFlow Matching-19.71±0.15-19.93±0.050.22

The repeated-evaluation statistics and performance differences have been added to Appendix G.2 of the uploaded revised manuscript.

W1-2: The encoder population behind each macro-average is not stated either, so differences of this size cannot be interpreted.

We apologize for the confusion caused by our description of the encoder population and aggregation procedure. Each family in each dataset contains 1213 independently trained encoders. Under unseen training configurations, the train/validation/test split is 973/121/119 per family, or 6811/847/833 across the seven families.

The seven families are CsiNet, CRNet, CLNet, TransNet, ResNet, MLP-AE, and Attention-CNN, totaling 8491 independent instances per dataset. Each independent training run contributes one checkpoint selected by validation performance; models from the same run and their derived data belong to the same task split. Unseen architectures are evaluated using seven leave-one-family-out folds, each testing all 1213 instances of the held-out family.

For each encoder, we accumulate reconstruction-error energy and channel energy over the query CSI and convert their ratio to dB to obtain its NMSE. We then average within each family and finally average equally over the seven families. Appendix G.1 of the uploaded revised manuscript gives the encoder counts and aggregation procedure.

W1-3: The unhedged superlative "lowest NMSE among generative methods in all 8 settings" nevertheless appears in the abstract, the results text and the conclusion.

We appreciate your suggestion to refine the wording of our conclusions. In the uploaded revised manuscript, the abstract, Section 5.2, and the conclusion all qualify the claim: FastAdapter achieves the lowest mean NMSE among the compared generative methods in all eight settings. The mean improvements of 0.17–2.17 dB over the strongest generative baseline in Table 1 support this statement.

AuthorsOpenReview ↗

Response to Reviewer joww W2 & Q2

W2-1: The contribution of the core mechanism is never isolated, and diffusion accounts for three quarters of the deployment cost. [...] The clean in-model control — the paper's own Gram–Raw condition encoder with a deterministic regression head under the same tokens, the same targets and the same training budget — appears nowhere,
Q2: Deterministic control. What NMSE does your own Gram–Raw condition encoder reach with a deterministic regression head under the same tokens, the same targets and the same training budget, and what is its deployment time against the 5.99 s pipeline? The limitations section names this experiment directly — why is it not in the ablation table? Related: was the HyperTransformer reimplementation deterministic?

Thank you for suggesting a deterministic control. MLP-L adapts faster, but its mean NMSE is 7.43–8.35 dB higher than FastAdapter's under unseen training configurations. Tables 2 and 3 compare accuracy and runtime.

  1. The MLP controls replace only the generator. The existing ablations test condition construction and ParamRealign. In the added control, MLP-S and MLP-L use the same Gram–Raw encoder, per-token context, training targets, and training budget as FastAdapter.
  2. Our HyperTransformer reimplementation is deterministic. Our reimplementation predicts the same Adapter for fixed model parameters and support codewords. All methods use comparable tuning budgets and validation-based model selection.
  3. DDIM sampling accounts for most of FastAdapter’s adaptation time. On COST2100/in, FastAdapter takes 2.92 s longer than MLP-L and achieves 8.35 dB lower mean NMSE. DDIM sampling accounts for 4.33±0.16 s of the 5.75±0.18 s total. Subsequent feedback reuses the Adapter.

Table 2. NMSE of deterministic predictors and FastAdapter under unseen training configurations (dB).

MethodCOST2100/inCOST2100/outWAIR-DDeepMIMO
MLP-S-2.16±0.15-1.18±0.27-1.00±0.22-0.43±0.31
MLP-L-13.16±0.16-7.34±0.14-12.83±0.15-13.67±0.23
HyperTransformer-16.15±0.07-9.71±0.15-17.04±0.13-17.51±0.14
FastAdapter-50-21.51±0.09-15.66±0.17-21.05±0.15-21.10±0.13

Table 3. Total adaptation time of deterministic predictors and FastAdapter under unseen training configurations (s).

MethodCOST2100/inCOST2100/outWAIR-DDeepMIMO
MLP-S1.71±0.121.72±0.131.82±0.121.76±0.15
MLP-L2.83±0.072.86±0.133.14±0.203.08±0.15
HyperTransformer2.91±0.812.86±0.773.51±0.923.14±0.89
FastAdapter-505.75±0.186.08±0.156.16±0.186.06±0.29

In the uploaded revised manuscript, Appendix G.3 describes the controls, and Appendices G.2 and G.5 report NMSE and runtime, respectively, for both protocols.

W2-2: Target preparation sits in the same position. The choice the paper argues for in prose — refining a candidate codeword through the frozen decoder rather than regressing pointwise to a reference encoder's codewords — is not ablated, and the target Adapter's own NMSE is reported nowhere, so the 0.96–2.65 dB gap to the supervised oracle cannot be attributed between preparation and generation.

Thank you for suggesting this comparison. We added a target-preparation ablation. Under unseen training configurations, target refinement lowers fitted Adapter NMSE by 0.09–0.96 dB. Generated Adapters trail Individual Adapters by 1.04–1.83 dB.

  1. Refined targets improve the fitted Adapters. We fix the Adapter architecture and training settings and change only the target from reference codewords E0(H)E_0(H) to refined codewords z~τ\tilde z_\tau. Fitted Adapter NMSE improves on all four datasets, from -22.76±0.21 to -23.18±0.13 dB on COST2100/in (0.42 dB).
  2. Individually fitted Adapters achieve lower NMSE than generated Adapters. Target Adapters fitted to training encoders with refined codewords supervise the generator. On test encoders, we fit Individual Adapters separately with paired CSI using the same procedure and compare them with the generated Adapters.

Table 4. NMSE results for target-preparation and parameter-generation diagnostics under unseen training configurations (dB).

ObjectCOST2100/inCOST2100/outWAIR-DDeepMIMO
Reference-codeword decoding-24.12-17.63-23.93-23.35
Source-model reconstruction-23.47-17.03-23.18-22.78
Refined-codeword decoding-23.41±0.05-16.94±0.06-23.10±0.04-22.71±0.05
Reference-target Adapter-22.76±0.21-15.88±0.20-22.36±0.17-22.05±0.21
Individual Adapter-23.18±0.13-16.84±0.16-22.88±0.18-22.14±0.07
FastAdapter-50-21.51±0.09-15.66±0.17-21.05±0.15-21.10±0.13

In Table 4, the fixed decoder reconstructs CSI from reference and refined codewords. Decoding refined codewords approaches the source model’s reconstruction with its own decoder. The corresponding results for unseen architectures appear in Appendix G.4 of the uploaded revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer joww W3 & Q3

W3: Smaller things, not driving my recommendation: "training configuration", the split unit of the headline protocol, is never defined; [...] all evaluation uses unquantized codewords drawn from the encoder's training split, while the appendix describes deployment as forming the support set from "previously received feedback codewords".

Thank you for checking these details.

  1. Training configuration. In preliminary experiments within each dataset and family, the training seed had the largest effect. We vary it while fixing the batch size, peak learning rate, learning-rate schedule, and optimizer. We hold out new instances within a family or one entire family (uploaded revised manuscript, Sections 3 and 5.1 and Appendix G.1).
  2. Figure 4 uses 973 training Adapters. We corrected 946 to 973 in the uploaded revised manuscript (Appendix A). The 1213 COST2100/in TransNet Adapters split 973/121/119 across train/validation/test.
  3. Interface dimensions and training hyperparameters. Tables 5 and 6 summarize the settings in Appendix G.1 of the uploaded revised manuscript.
  4. Online Affine and Affine Adapter denote the same baseline, as clarified in Section 5.1 of the uploaded revised manuscript.
  5. Independent support gives similar NMSE. Codewords from encoder-training CSI emulate feedback already received at the base station. The new support pool excludes encoder-training and checkpoint-selection CSI and is disjoint from the query set. Switching pools changes mean NMSE by at most 0.09 dB across the four datasets under unseen training configurations (Table 7). The codewords are unquantized. The 1/4 ratio refers to codeword length.

Table 5. Input dimensions, codeword dimensions, and compression ratios in the main experiments.

SettingCOST2100/inCOST2100/outWAIR-DDeepMIMO
Real-valued CSI input dimensions2×32×322×32×322×64×642×32×64
Codeword dimension C51251220481024
Compression ratio CR1/41/41/41/4

Table 6. Key settings in the main experiments.

SettingValue
Number of residual FFN blocks in the Adapter4
FFN hidden width512
CFG guidance strength1.5
InfoNCE temperature0.07
Target Adapter training batch size32
Batch size for contrastive pretraining and generator training32 encoder tasks
Context width512

Table 7. NMSE with different support CSI sources under unseen training configurations (dB).

DatasetTraining-CSI supportIndependent-CSI support
COST2100/in-21.51±0.09-21.46±0.10
COST2100/out-15.66±0.17-15.64±0.18
WAIR-D-21.05±0.15-20.98±0.17
DeepMIMO-21.10±0.13-21.01±0.21

In the uploaded revised manuscript, Appendix G.6 covers unseen architectures and Section 6 discusses the limitation of unquantized feedback.

Q3: On which split were the 55 grid cells computed? [...] Related: why is λ_align fixed at 1 and never searched, why does λ_ref = 2 sit at the top of the search range, and were the WAIR-D and DeepMIMO weights transferred from COST2100/in without validation? [...]

Thank you for asking how we selected the loss weights.

  1. Validation selects the weights; the 55-cell grid reports test NMSE. For each candidate, we fitted target Adapters and trained the generator. We selected weights by the reconstruction NMSE of generated Adapters on validation encoders, then evaluated on independent test encoders. In the uploaded revised manuscript, Appendix B and Section 5.2 report the same test NMSE for the selected weights. Table 8 also gives the validation optimum.

Table 8. Validation selection and test evaluation of generated Adapters, NMSE (dB).

ProtocolBest validation-grid NMSETest mean±std
Unseen training configurations-21.82-21.51±0.09
Unseen architectures-21.32-21.16±0.14
  1. The validation optimum remains at λref=2\lambda_{\mathrm{ref}}=2 after expanding the search. We fixed λalign=1\lambda_{\mathrm{align}}=1 as the reference scale for the other weights. The extended grid includes λref=2.5\lambda_{\mathrm{ref}}=2.5 and 33, with 77 configurations per protocol. Both protocols reach their lowest validation NMSE at (λalign,λrecon,λref)=(1,1000,2)(\lambda_{\mathrm{align}},\lambda_{\mathrm{recon}},\lambda_{\mathrm{ref}})=(1,1000,2) (Table 9). WAIR-D and DeepMIMO reuse the COST2100/in weights without a separate search. After fitting all target Adapters, we split encoder tasks using seed 0. All generators use the same training-set target Adapter parameters for supervision.

Table 9. For each fixed λref\lambda_{\mathrm{ref}}, the best λrecon\lambda_{\mathrm{recon}} and validation NMSE (dB).

λref\lambda_{\mathrm{ref}}Unseen training configurations: best λrecon\lambda_{\mathrm{recon}}Validation NMSEUnseen architectures: best λrecon\lambda_{\mathrm{recon}}Validation NMSE
1.51000-21.301200-21.28
21000-21.821000-21.32
2.51000-21.531000-21.12
31000-21.221000-20.98

Both expanded grids appear in Appendix G.7 of the uploaded revised manuscript.

Thank you for the thorough rebuttal and for diligently addressing the experimental gaps in the revised manuscript. Overall, the problem formulation is clean and the paper is well-deserving of acceptance. However, given the modest margins over the strongest generative baselines in several settings (0.17–0.27 dB), I will maintain my rating of Weak Acceptance.

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

This paper investigates the "codeword drift" problem in Massive MIMO CSI feedback systems, where independently trained or updated UE-side encoders produce representations incompatible with a fixed BS-side decoder. The authors propose FastAdapter, a framework that generates a complete BS-side Adapter from unlabeled codewords (support set) uploaded by the new encoder, achieving cross-encoder alignment without gradient updates or paired CSI supervision. Experiments on COST2100, WAIR-D, and DeepMIMO cover unseen training configurations and unseen architectures generalization protocols, and demonstrate that FastAdapter outperforms all generative baselines in NMSE. Further efficiency experiments and analytical ablations validate the effectiveness of FastAdapter.

Critical strengths
  1. The paper presents substantial content with a well-motivated problem formulation. Two core challenges are clearly identified: (a) unlabeled codewords entangle channel content with encoder-specific representations; (b) independently trained Adapters exhibit permutation symmetries that produce inconsistent parameter targets. The proposed module designs address these two challenges in a targeted manner.
  2. The experimental evaluation is comprehensive. It includes not only supervised methods but also four generative baselines and four non-generative baselines, providing a thorough performance assessment across multiple dimensions.
  3. The ablation studies are thorough. Ablations are conducted for ParamRealign, Gram/Raw branches, position IDs, global condition, and contrastive pretraining, covering both generalization protocols across 8 experimental settings. These ablations provide sufficient evidence to support the effectiveness of each component.
Critical weaknesses
  1. Although the paper employs a diffusion model for parameter generation, the justification for choosing diffusion over other generative paradigms is not fully convincing. The authors do not provide a clear argument or experimental evidence for why diffusion is necessary here, nor do they discuss whether alternative generative models (e.g., autoregressive generation, flow matching, or a strong deterministic regressor) could serve as adequate substitutes. The paper itself acknowledges in the Limitations section that "a deterministic generator trained with the same conditions would further clarify the contribution of diffusion," yet no such comparison is provided.。
  2. My main concern lies in the efficiency analysis. First, Table 3 reports adaptation times for FastAdapter, Individual Adapter, Online Affine, and Nearest Adapter, but omits the inference times of the four generative baselines (HyperTransformer, Cond P-Diff, D2NWG, RPG). Given that the slow sampling speed of diffusion models is a well-known limitation, and that the core objective of this paper is efficiency, comparing against other generative baselines in terms of runtime is essential. Second, the paper fixes DDIM sampling at 50 steps. What are the NMSE and generation times under different numbers of steps? Intuitively, more steps should yield better generation quality (lower NMSE) but longer latency. How should this trade-off be navigated? A sensitivity analysis over step counts would be informative. Third, the paper does not report GPU memory consumption. For practical deployment scenarios with limited storage resources, this is a critical reference metric.
Questions for authors
  1. Could the authors provide a formal justification or intuitive explanation for why diffusion is the appropriate generative paradigm for this task? Specifically, what properties of the Adapter parameter distribution necessitate iterative denoising rather than single-pass generation?
  2. Could the authors supplement the efficiency analysis with the following experiments: (i) inference/generation times of all four generative baselines; (ii) NMSE and total adaptation time under varying DDIM step counts (e.g., 10, 20, 50, 100); and (iii) peak GPU memory consumption during adaptation? Since the paper's central claim revolves around efficiency, these analyses are critical for assessing practical use.
5 comments
AuthorsOpenReview ↗

Response to Reviewer BvDn W1

Thank you for your careful review and constructive comments. We appreciate your positive assessment of our targeted module design and comprehensive evaluation. Below are our responses to your comments (your comments are highlighted in italics).

W1: Although the paper employs a diffusion model for parameter generation, the justification for choosing diffusion over other generative paradigms is not fully convincing. [...] whether alternative generative models (e.g., autoregressive generation, flow matching, or a strong deterministic regressor) could serve as adequate substitutes. [...]

We apologize for the confusion caused by our explanation of the choice of diffusion. Our main reasons for choosing conditional diffusion are the precision required for parameter generation, the observed properties of the parameter distribution, the need for conditional generation, and the performance advantages in controlled comparisons.

  1. Precision requirements of parameter generation. The quality of parameter generation is ultimately determined by the CSI reconstruction accuracy obtained with the complete generated Adapter and the frozen decoder. Generated parameters must therefore preserve the mapping jointly implemented by the weights and biases across layers. In image generation, early methods were mostly deterministic regression networks with an L2 objective; their outputs tend toward the conditional mean, and pixel errors affect only local parts of the image. An Adapter, however, acts on every codeword: a deviation in a single weight changes the transformation of all codewords, and the nonlinear decoder amplifies the error throughout the CSI reconstruction. Parameter generation therefore requires greater precision.
  2. Properties of the parameter distribution. In our preliminary target Adapter preparation experiments, we initialized parameters from a Gaussian distribution. After training, we inspected the distribution of parameter values and observed that it remained approximately Gaussian. Per-tensor standardization brings different weights and biases to a common scale, allowing the generator to learn to recover these parameters under the same Gaussian noise schedule.
  3. Parameter generation requires continued use of the condition. At every denoising step, FastAdapter reads both the current parameter state and the full per-token context CτC_\tau to update the complete parameter set. The encoder information provided by the support codewords thus continues to guide adjustments as the parameters are progressively recovered. The added MLP controls also use the full CτC_\tau and the same training budget as FastAdapter, but output parameters in a single forward pass, allowing us to compare direct prediction with iterative recovery under the full condition.
  4. Controlled comparisons support the accuracy advantage of diffusion. HyperTransformer is already a deterministic parameter predictor; we further add MLP-S, MLP-L, a Transformer-based autoregressive generator, and Flow Matching. Table 1 shows that FastAdapter achieves lower mean NMSE under unseen training configurations: 7.43–8.35 dB lower than MLP-L, whose capacity is close to FastAdapter's; 3.59–5.95 dB lower than HyperTransformer; and 8.24–8.78 dB and 0.62–2.19 dB lower than the autoregressive generator and Flow Matching, respectively.

For this revision, we repeated the evaluation with seeds 42, 2026, and 3407 and report mean±std.

Table 1. NMSE of parameter-generation methods under unseen training configurations (dB, lower is better).

MethodCOST2100/inCOST2100/outWAIR-DDeepMIMO
MLP-S-2.16±0.15-1.18±0.27-1.00±0.22-0.43±0.31
MLP-L-13.16±0.16-7.34±0.14-12.83±0.15-13.67±0.23
HyperTransformer-16.15±0.07-9.71±0.15-17.04±0.13-17.51±0.14
Autoregressive-13.23±0.15-7.09±0.13-12.27±0.20-12.86±0.19
Flow Matching-20.31±0.22-13.47±0.19-20.43±0.17-20.36±0.13
FastAdapter-21.51±0.09-15.66±0.17-21.05±0.15-21.10±0.13

These results support conditional diffusion for parameter generation. In the uploaded revised manuscript, Appendix G.2 reports reconstruction accuracy under both protocols, and Appendix G.3 describes the new generator comparisons.

AuthorsOpenReview ↗

Response to Reviewer BvDn W2-1/Q2-1 & W2-2/Q2-2

W2-1: My main concern lies in the efficiency analysis. First, Table 3 reports adaptation times for FastAdapter, Individual Adapter, Online Affine, and Nearest Adapter, but omits the inference times of the four generative baselines (HyperTransformer, Cond P-Diff, D2NWG, RPG). [...]

Q2-1: Could the authors supplement the efficiency analysis with the following experiments: (i) inference/generation times of all four generative baselines;

Thank you for highlighting adaptation efficiency. Under unseen training configurations, FastAdapter's total adaptation time is 5.75–6.16 s, shorter than the 42.45–120.72 s required by Cond P-Diff, D2NWG, and RPG. Deterministic predictors take less time, while FastAdapter achieves higher reconstruction accuracy. We timed all generators for this revision. Table 2 reports timings for unseen training configurations. In the uploaded revised manuscript, Appendix G.5 covers both protocols.

  1. Direct predictors adapt faster, while FastAdapter achieves lower reconstruction error. On COST2100/in under unseen training configurations, FastAdapter takes 2.92 s longer than MLP-L while reducing mean NMSE by 8.35 dB. Across the four datasets, HyperTransformer takes 2.86–3.51 s, with mean NMSE 3.59–5.95 dB higher than FastAdapter.
  2. Compared with other iterative generative baselines, FastAdapter adapts faster. Under unseen training configurations, Cond P-Diff, D2NWG, and RPG take longer than FastAdapter on all four datasets. The added autoregressive generator and Flow Matching take 14.28–18.74 s and 32.14–39.83 s, respectively. Together with Table 1, these results show that FastAdapter has both accuracy and runtime advantages over these two methods as well.

Table 2. Total adaptation time of generators under unseen training configurations (s).

MethodCOST2100/inCOST2100/outWAIR-DDeepMIMO
MLP-S1.71±0.121.72±0.131.82±0.121.76±0.15
MLP-L2.83±0.072.86±0.133.14±0.203.08±0.15
HyperTransformer2.91±0.812.86±0.773.51±0.923.14±0.89
Cond P-Diff45.98±5.4142.45±4.3650.09±4.9246.79±4.54
D2NWG92.98±7.4489.04±8.0388.76±7.4691.09±7.45
RPG103.92±7.8999.53±8.41120.72±10.49115.64±9.38
Autoregressive15.63±1.4714.28±1.8318.74±2.2917.52±1.94
Flow Matching34.72±3.9132.14±3.4739.83±4.2837.56±4.01
FastAdapter-505.75±0.186.08±0.156.16±0.186.06±0.29

W2-2: Second, the paper fixes DDIM sampling at 50 steps. What are the NMSE and generation times under different numbers of steps? Intuitively, more steps should yield better generation quality (lower NMSE) but longer latency. How should this trade-off be navigated? A sensitivity analysis over step counts would be informative.

Q2-2: (ii) NMSE and total adaptation time under varying DDIM step counts (e.g., 10, 20, 50, 100);

We appreciate your question about the tradeoff between sampling steps and adaptation efficiency. We choose 50 steps because the accuracy gains diminish markedly beyond this point. In the new step-count experiment, we evaluated 1, 5, 10, 20, 50, and 100 steps. Under unseen training configurations, increasing from 50 to 100 steps improves mean NMSE by only 0.01–0.05 dB while adding 3.23–4.48 s. On COST2100/in, increasing from 20 to 50 steps improves mean NMSE by 4.20 dB, whereas increasing from 50 to 100 steps improves it by only 0.04 dB. Tables 3 and 4 report the full step-count comparison on COST2100/in and the 50-/100-step results across all four datasets, respectively.

Table 3. NMSE and total adaptation time at different DDIM step counts on COST2100/in under unseen training configurations.

DDIM stepsNMSE (dB)Time (s)
1-1.87±0.291.74±0.05
5-6.37±0.261.92±0.09
10-11.37±0.212.37±0.14
20-17.31±0.223.21±0.23
50-21.51±0.095.75±0.18
100-21.55±0.149.22±0.48

Table 4. Comparison of 50-step and 100-step DDIM across four datasets under unseen training configurations.

Dataset50-step NMSE (dB)100-step NMSE (dB)50-step time (s)100-step time (s)
COST2100/in-21.51±0.09-21.55±0.145.75±0.189.22±0.48
COST2100/out-15.66±0.17-15.71±0.126.08±0.159.31±0.37
WAIR-D-21.05±0.15-21.09±0.116.16±0.1810.64±0.39
DeepMIMO-21.10±0.13-21.11±0.136.06±0.299.93±0.41

The step-count results for unseen architectures appear in Appendix G.5 of the uploaded revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer BvDn W2-3/Q2-3 & Q1

W2-3: Third, the paper does not report GPU memory consumption. For practical deployment scenarios with limited storage resources, this is a critical reference metric. 

Q2-3: and (iii) peak GPU memory consumption during adaptation? Since the paper's central claim revolves around efficiency, these analyses are critical for assessing practical use.

Thank you for suggesting that we report deployment-time GPU memory usage. Under unseen training configurations, FastAdapter's peak GPU memory usage is 0.50–3.63 GB. Table 5 reports our new peak-memory measurements. On COST2100/in, FastAdapter uses more GPU memory than MLP-S, MLP-L, and HyperTransformer, and less than the other generative baselines. The ranking varies across datasets: for example, on WAIR-D under unseen training configurations, MLP-L and RPG use 3.28±0.41 GB and 3.21±0.26 GB, respectively, less than FastAdapter's 3.63±0.38 GB.

Table 5. Peak GPU memory usage of generators under unseen training configurations (GB).

MethodCOST2100/inCOST2100/outWAIR-DDeepMIMO
MLP-S0.23±0.030.24±0.021.08±0.120.60±0.05
MLP-L0.41±0.100.41±0.083.28±0.411.11±0.15
HyperTransformer0.43±0.130.44±0.103.67±0.390.98±0.14
Cond P-Diff0.95±0.210.95±0.184.32±0.422.03±0.32
D2NWG1.70±0.271.72±0.213.60±0.412.08±0.27
RPG0.82±0.190.80±0.213.21±0.261.43±0.20
Autoregressive0.89±0.170.87±0.154.05±0.371.78±0.26
Flow Matching0.90±0.190.89±0.183.92±0.401.88±0.30
FastAdapter0.51±0.160.50±0.133.63±0.381.54±0.31

Peak memory under unseen architectures is reported in Appendix G.5 of the uploaded revised manuscript.

Q1: Could the authors provide a formal justification or intuitive explanation for why diffusion is the appropriate generative paradigm for this task? Specifically, what properties of the Adapter parameter distribution necessitate iterative denoising rather than single-pass generation?

Thank you for asking for further clarification of iterative denoising. Regarding the choice of diffusion, as explained in our response to W1, we considered the precision required for parameter generation, the observed properties of the parameter distribution, how the condition is used, and the measured performance. The comparisons with deterministic predictors, the autoregressive generator, and Flow Matching in Table 1 support this choice.

The reason for using iterative denoising rather than single-pass generation can be explained through how the condition is used. As discussed in W1, parameter generation requires continued use of the condition. Specifically, at each step, FastAdapter feeds the current parameter state XtX_t and the full per-token context CτC_\tau into the denoising network, predicts the noise ϵψ(Xt,t,Cτ)\epsilon_\psi(X_t,t,C_\tau), and updates the parameters accordingly. As the parameter state changes, the same support condition continues to guide subsequent adjustments, forming the final Adapter through repeated joint updates of parameters across layers.

For comparison, let Θ\Theta denote the realigned target Adapter parameters and cc a given conditional context. When a direct predictor is trained with squared parameter error, the prediction that minimizes the expected loss is the conditional mean:

f∗(c)=argmin⁡uE[∥Θ−u∥22∣Cτ=c]=E[Θ∣Cτ=c].f^{\ast}(c)=\underset{u}{\operatorname{argmin}} \mathbb{E}[\lVert\Theta-u\rVert_2^2\mid C_{\tau}=c]=\mathbb{E}[\Theta\mid C_{\tau}=c].

Let gτ,H(θ)=D0(Aθ(Eτ(H)))g_{\tau,H}(\theta)=D_0(A_\theta(E_\tau(H))) denote the CSI reconstructed through the Adapter. Because gτ,Hg_{\tau,H} is nonlinear in the parameters θ\theta, minimizing squared parameter error does not guarantee minimizing CSI reconstruction error. Conditional denoising instead recovers the complete parameter set through the multi-step updates described above; its benefit is evaluated through the direct-predictor controls and step-count experiments.

Under unseen training configurations with the full condition, MLP-L's mean NMSE is 7.43–8.35 dB higher than FastAdapter's (Table 1). Within the same diffusion model, increasing from 1 to 50 steps on COST2100/in improves NMSE from −1.87±0.29 to −21.51±0.09 dB under unseen training configurations (Table 3). The direct-predictor controls and sampling-step experiments jointly support the iterative conditional generation scheme used in this work. The mechanism explanation and step-count evidence have been added to Appendices G.3 and G.5 of the uploaded revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.