acceptodds
Under review as a conference paper at ICLR 2027

What Can WiFi Say? Language-Aligned Representation Learning for WiFi Signals

Abstract

Understanding everyday activity could support smarter homes, personal assistants, and remote care. Camera-based monitoring provides rich observations but raises privacy concerns, while WiFi offers a privacy-aware alternative. We therefore ask: What can WiFi say in natural language? With language, WiFi could describe movement, retrieve related activities, and answer questions about what happened. We introduce SignalLang, a unified framework that aligns WiFi channel state information (CSI) with motion-focused language. Its captions retain posture, body-part involvement, motion attributes, and temporal structure while excluding unsupported visual details. A transformer over subcarrier-time tokens preserves the link, frequency, and temporal structure of CSI, while contrastive, matching, generative, and auxiliary activity objectives learn a shared WiFi-text representation. Class-aware negative masking prevents recordings of the same activity from being treated as negatives. A single backbone supports caption generation, bidirectional retrieval, and WiFi question answering. On XRF55 activities represented during training, SignalLang improves over an adapted WiFi2Cap baseline by 19.9 BLEU-4 points, 51.8 percentage points in WiFi-to-text retrieval R@1, and 25.3 percentage points in WiFi question answering accuracy. It also shows measurable transfer to held-out activity classes in retrieval and question answering.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
WiFi Sensing, Multimodal Learning, Language-Aligned Representation Learning

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 2 reviews, above 85% of the others.
Reviewer_fjPFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies WiFi-language representation learning and cross-modal retrieval in a WiFi sensing setting. The proposed SignalLang framework combines a WiFi encoder operating on subcarrier–time patches with contrastive, matching, language-modeling, and auxiliary activity-classification objectives. It also introduces motion-focused caption supervision and class-aware negative masking. SignalLang supports caption generation, bidirectional retrieval, and WiFi question answering. Experiments show substantial improvements over an amplitude-only WiFi2Cap adaptation on seen XRF55 activities.

Critical strengths
  • The study is well motivated to align WiFi signals with language descriptions.

  • The supervision and language description generated are sensor-aware. The paper identifies an important mismatch between video-derived descriptions and information supported by WiFi CSI. Its motion-focused supervision aims to retain motion semantics while excluding unsupported visual details. This is a useful principle for transferring supervision between sensors with different observation capabilities.

  • The proposed method covers caption generation, retrieval, and WiQA. The study also distinguishes activity-level retrieval from instance-specific matching, its hallucination proxy from verified factual errors, and framework portability from zero-shot checkpoint generalization.

Critical weaknesses
  • The retrieval results support a learnable association between WiFi and activity semantics. However, retrieval relevance is defined by activity labels, and caption evaluation includes references from the same activity class. These protocols do not establish whether the model recovers recording-specific body-part movements, directions, repetitions, or temporal structure. The FullCap versus SimpleCap comparison is useful, but it changes both descriptive content and linguistic diversity. The evaluation of WiQA also lacks reported question-only and signal-replacement controls. Consequently, activity recognition followed by class-conditioned language generation remains a plausible explanation for a substantial part of the performance. The support for fine-grained WiFi–language understanding is limited.

  • Most training processes follow commonly used multimodal learning approaches. Subcarrier–time tokenization provides a reasonable CSI-specific adaptation, but the main ablations do not directly test this design against alternative encoders under matched conditions. They therefore do not establish whether the gains come from the proposed signal representation, additional activity supervision, or the combined training objectives.

  • Most baseline comparisons and component ablations on XRF55. XRF V2 was used mainly to demonstrate portability after training, without evaluating matched competing methods or corresponding component validation as used on XRF55.

Questions for authors

Please refer to the points listed in Critical Weaknesses.

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer fjPF

We thank the reviewer for the constructive feedback and for recognizing the motivation and contributions of our work. We have conducted additional experiments to address the concerns.

1. Activity-Level Semantics and WiQA Controls

We acknowledge that our retrieval protocol evaluates activity-level semantic alignment rather than instance-specific matching. Recordings of the same activity in XRF55 often share similar motion patterns and semantically compatible descriptions, making exact-pair retrieval potentially ambiguous. Therefore, our current retrieval results support activity-level motion-language alignment, but do not establish discrimination among recordings of the same activity.

To assess whether WiQA predictions depend on valid CSI inputs rather than linguistic priors, we evaluated three additional settings: question-only (without CSI), shuffled-CSI (CSI from another recording), and zero-CSI (all-zero input). SignalLang achieves 72.7% overall accuracy, compared with 39.7%, 36.2%, and 43.5%, respectively.

For body-part questions, SignalLang achieves 84.4% vs. 52.8% under shuffled CSI on seen activities and 61.1% vs. 37.8% on unseen activities. On seen motion-attribute questions, accuracy reaches 69.5% vs. 28.2% under shuffled CSI. These results show that valid CSI provides substantial information beyond linguistic priors, supporting CSI-dependent activity-level motion understanding. However, they do not establish recording-specific motion discrimination, which remains an important limitation of the current evaluation.

2. Additional Ablation Studies

We conducted additional experiments to evaluate the effects of different WiFi encoders and key training components. Specifically, we evaluated a randomly initialized Transformer encoder and a 1-D ResNet encoder, alongside variants without auxiliary activity supervision or class-aware negative masking. The results are summarized below:

VariantSetWi2TR R@1T2WiR R@1B@4METEORROUGE-LCIDErLLM Cls.@1WiQA Acc.
WTC + WTM + LM only (no ACT, no mask)Seen39.742.440.158.455.9114.353.379.9
Unseen23.325.211.934.536.113.824.239.1
w/o auxiliary activity lossSeen40.061.440.159.056.3113.445.780.6
Unseen29.129.811.334.836.313.222.539.3
Transformer, random initSeen64.985.243.361.458.6126.857.381.7
Unseen28.832.612.235.537.114.325.939.4
1-D ResNet, random initSeen84.395.751.167.664.4160.666.085.8
Unseen40.551.014.536.838.617.126.842.0
SignalLang (full; Transformer, MAE init)Seen71.589.843.861.758.8129.860.680.8
Unseen32.735.611.535.036.514.124.846.2

Note: The plain BLIP-style baseline removes both auxiliary activity supervision and class-aware negative masking.

The results show that the plain BLIP-style baseline, retaining only WTC, WTM, and LM, performs substantially worse than full SignalLang. Its seen Wi2TR/T2WiR R@1 is 39.7%/42.4%, compared with 71.5%/89.8% for full SignalLang, while multi-reference CIDEr is 114.3 versus 129.8. Together with the individual ablations in our original submission, these results support the contributions of auxiliary activity supervision and class-aware negative masking.

The 1-D ResNet also outperforms the Transformer on XRF55, potentially because its convolutional inductive bias is better suited to the limited diversity of scripted motions in this dataset. Transformer-based encoders may benefit more from larger and more diverse training datasets, which remains to be investigated. These results indicate that the WiFi–text learning framework can accommodate different encoder architectures, while its performance depends on the encoder design.

AuthorsOpenReview ↗

Author Response to Reviewer fjPF (Continued)

3. Additional Evaluation on XRF V2

We extended our XRF V2 experiments with a matched WiFi2Cap* baseline and additional ablation studies. Specifically, we evaluated the effects of auxiliary activity supervision, class-aware negative masking, and MAE initialization. The results are summarized below:

VariantWi2TR R@1T2WiR R@1B@4METEORROUGE-LCIDErLLM Cls.@1
w/o auxiliary activity loss42.948.742.155.464.5344.937.7
w/o class-aware negative masking39.939.442.755.964.5336.737.9
w/o MAE initialization47.8 / 44.573.8 / 74.244.8 / 42.158.7 / 59.566.4 / 64.7363.7 / 329.842.4 / 36.9
WiFi2Cap*20.644.033.348.756.4205.417.6
SignalLang (full)45.4 / 45.670.9 / 54.041.5 / 41.754.8 / 58.563.7 / 64.4326.6 / 317.336.2 / 36.6

Note: Slash-separated values represent results from two different random seeds, reported in the same order across metrics. Rows with a single value report one experimental run.

SignalLang outperforms WiFi2Cap* across all reported retrieval and captioning metrics on XRF V2. Removing auxiliary activity supervision or class-aware negative masking reduces bidirectional retrieval performance, further supporting the effectiveness of these components on a second dataset.

In contrast, MAE initialization does not consistently improve performance on XRF V2. This suggests that the benefits of MAE pretraining depend on the dataset and training setting, rather than being essential to the proposed WiFi–text alignment framework. Overall, these results demonstrate the portability of SignalLang after retraining on a different WiFi acquisition setting, while highlighting the contributions of its activity-aware alignment components.

Thank you for the response. The new WiQA controls and XRF V2 comparisons address part of my concerns. However, the concerns about recording-specific fine-grained understanding remain. Moreover, the encoder comparisons still do not show a clear advantage of the proposed WiFi encoding design. Nonetheless, these limitations do not undermine my rating.

Thanks for your reply. The additional experiments are sufficient; the reply is very good, and my concern is addressed. I will raise the point to 3.

Reviewer_Lz4wOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a framework that aligns WiFi Channel State Information with motion-related natural language. It provides a unified approach supporting caption generation, WiFi-text retrieval, and WiFi-based Question Answering. The methodology incorporates MAE pre-training, multi-objective WiFi-text alignment, and motion-focused caption supervision. Experimental results show that it outperforms other baselines.

Critical strengths
  1. The research topic is interesting. the authors map WiFi data into a natural language semantic space instead of only pursuing the traditional classification task. This allows a single backbone model to support multiple tasks related to WiFi signal.
  2. The design of motion-focused supervision is interesting. The authors recognize that much of the information found in standard video captions—such as clothing color, scene details, or object appearance that cannot be perceived by WiFi. And also it use experiments to show the design is good.
Critical weaknesses

1: Insufficient Baseline Comparisons The primary experiments are compared almost exclusively against WiFi2Cap*—a version modified by the authors to handle amplitude-only inputs. Related works like WirelessSenseLLM and HoloLLM are not compared. More baselines need to be added to show the effectiveness of the proposed algorithm. 2: Insufficient Ablation Studies The paper provides specific ablation studies for some of the component of the design but the many core components are still not studied. WTC and WTM are not ablated individually. Consequently, it is currently hard to determine the individual contribution of each core module. 3: Lack of Comparison of Model Complexity and Computational Overhead The paper does not report the number of parameters or FLOPs for either SignalLang or WiFi2Cap*. SignalLang has several additional module and it is difficult to know whether the performance gains come from the increased model capacity or higher computational costs. It is recommended that the authors provide a comparison of parameter counts and computational overhead to ensure a fair evaluation.

Questions for authors

From the experiments we know that the performance of the model on unseen activity is not that well , that do you think is the main reason for this? Do you have any idea how to prevent this from happening.

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer Lz4w

We thank the reviewer for the constructive feedback and appreciate the time and effort spent on reviewing our paper. We have conducted additional experiments to address the concerns.

1. Additional Baseline Comparisons

We agree that broader baseline comparisons would strengthen the evaluation. Our original comparison with WiFi2Cap* uses the same CSI amplitude inputs and data splits as SignalLang. Since the preprocessed XRF55 inputs used in our experiments retain only CSI amplitude, we removed WiFi2Cap's phase encoder and amplitude–phase fusion module to ensure a matched-input comparison.

We have additionally evaluated WiFi2Cap* on XRF V2 under a matched protocol. SignalLang achieves 45.4% vs. 20.6% WiFi-to-text R@1, 70.9% vs. 44.0% text-to-WiFi R@1, and 41.5 vs. 33.3 BLEU-4, demonstrating improvements on a second dataset.

We also investigated another baseline, HoloLLM. For a fair comparison, we followed the authors' released two-stage pipeline using our XRF55 split, excluding the 10 held-out activities from training. In Stage 1, we trained the WiFi encoder through supervised activity classification, achieving 92.6% accuracy on seen activities. In Stage 2, we trained the LLaMA2-7B-based captioning and QA model using our annotations and the authors' released implementation and hyperparameters. The results are shown below:

MetricHoloLLM (LLaMA2-7B)SignalLang
WiQA accuracy, seen / unseen (%)76.7 / 39.480.8 / 46.2
WiQA blind-hard subset (%)65.170.1
Caption LLM-judge, seen / unseen (%)27.7 / 18.660.6 / 24.8
Caption CIDEr (multi-ref.), seen / unseen101.8 / 15.4129.8 / 14.1
Language-model parameters7B137M

Note: The blind-hard subset consists of WiQA questions that the question-only baseline answered incorrectly. Both models are evaluated on the same subset.

Both models are evaluated on the same XRF55 test recordings using identical caption annotations. Under the matched XRF55 protocol, SignalLang achieves higher seen caption LLM-judge accuracy (60.6% vs. 27.7%), multi-reference CIDEr (129.8 vs. 101.8), and seen/unseen WiQA accuracy (80.8%/46.2% vs. 76.7%/39.4%), despite using a substantially smaller language model (137M vs. 7B parameters). Performance on held-out captioning remains mixed.

WirelessSenseLLM focuses on zero-shot understanding of continuous, unsegmented CSI using video-supervised alignment and a large language model. In contrast, SignalLang learns from segmented, amplitude-only CSI with offline video-derived text supervision. Adapting WirelessSenseLLM to our setting would require substantial modifications to its input and training pipeline. We therefore prioritize more directly comparable baselines.

AuthorsOpenReview ↗

Author Response to Reviewer Lz4w (Continued)

2. Additional Ablation Studies

We agree that the individual contributions of WTC and WTM should be evaluated. We have conducted additional ablations by removing each objective individually. We also evaluated a simplified variant retaining only WTC, WTM, and LM, without auxiliary activity supervision or class-aware negative masking, to evaluate the effect of removing both components simultaneously. All experiments follow the same training schedule as the original submission to ensure a fair comparison. The results are summarized below:

VariantSeen R@1 (W/T)Held-out R@1 (W/T)Seen CIDErLLM-judge seenLLM-judge held-outWiQA overall
SignalLang (full)71.5 / 89.832.7 / 35.6129.860.624.872.7
w/o WTC20.0 / 38.814.9 / 22.5130.560.424.173.3
w/o WTM50.4 / 71.625.5 / 26.7131.660.825.073.6
WTC + WTM + LM only39.7 / 42.423.3 / 25.2114.353.324.272.5
w/o MAE initialization64.9 / 85.228.8 / 32.6126.857.325.974.0

Note: W/T denotes WiFi-to-text / text-to-WiFi retrieval. CIDEr uses multi-reference evaluation. WiQA overall refers to the additional evaluation.

The results show that removing either WTC or WTM substantially degrades bidirectional retrieval performance, while having relatively minor effects on captioning and WiQA. The variant using only WTC, WTM, and LM also performs considerably worse in retrieval and captioning than full SignalLang, highlighting the benefits of auxiliary activity supervision and class-aware negative masking. In contrast, removing MAE initialization leads to relatively small and mixed changes, suggesting a less consistent contribution across tasks.

3. Model Complexity and Computational Overhead

We additionally measured the parameter counts and computational costs of SignalLang and WiFi2Cap*. The results are summarized below:

MetricSignalLangWiFi2Cap*
Trainable parameters165.4M (8.6M WiFi + 127.8M BERT + 29.0M decoder)925.4M (906.9M prefix MLP alone)
Parameters used for captioning137.1M1,049.8M
WiFi encoder GFLOPs / sample0.521.76
Caption generation GFLOPs / sample (as implemented)66.9 (greedy) / 252.6 (beam-3)85.1 (greedy)
Caption generation GFLOPs / sample (with KV cache)11.9 (greedy) / 39.3 (beam-3)8.3 (greedy)
Retrieval GFLOPs / WiFi query0.53 (ITC) / 806 (+ITM, top-128)1.77
Retrieval GFLOPs / text query5.13 (ITC) / 811 (+ITM)5.97
Training (A100-hours)13.8 + 0.6 (MAE)8.4

SignalLang uses substantially fewer trainable parameters than WiFi2Cap* (165.4M vs. 925.4M) and requires less computation for WiFi encoding (0.52 vs. 1.76 GFLOPs/sample). However, its training cost is higher, and ITM-based reranking introduces substantial additional retrieval overhead. Caption generation costs also vary with the decoding configuration and caching strategy. These results highlight SignalLang's advantages in model size and WiFi encoding efficiency, alongside the computational trade-offs of the full framework.

4. Performance on Unseen Activities

The 10 unseen activity classes are excluded from all training stages, requiring SignalLang to generalize learned WiFi–language alignment to previously unseen motions. We believe the performance gap mainly stems from two factors. First, WiFi CSI provides indirect measurements of human motion, making it difficult to distinguish activities with similar movement patterns. Second, XRF55 consists primarily of scripted repetitions of predefined activities, limiting the diversity of motion patterns available during training. Potential improvements include training on more diverse datasets with continuous and compositional motions, leveraging richer CSI measurements with wider bandwidth, and improving cross-domain robustness. We acknowledge that generalization to unseen activities remains challenging and will clarify this limitation in the revised paper.

Comments on the paper

3 comments

We thank the reviewer for the detailed and constructive feedback and appreciate the time and effort spent on reviewing our paper. We have conducted additional experiments and analyses to address the concerns.

1. Technical Contributions and Baseline Comparisons

While the individual multimodal learning objectives build upon established techniques, SignalLang addresses WiFi-specific challenges through a unified framework combining motion-focused supervision, structure-preserving CSI encoding, and activity-aware alignment. We use Qwen3.6-35B to generate motion-focused captions from synchronized videos offline, without incorporating visual features into model training or inference. Our CSI encoder preserves link, subcarrier, and temporal structure, while auxiliary activity supervision and class-aware negative masking retain activity-discriminative information and avoid same-class false negatives.

These designs distinguish SignalLang from prior wireless–language approaches in both learning framework and task setting. WiFi2Cap transfers video–text teacher representations to CSI for caption generation, while HoloLLM conditions a large language model on sensor embeddings. In contrast, SignalLang learns directly from WiFi–text pairs without visual-feature distillation or a large language model in its backbone. WirelessSenseLLM focuses on continuous, potentially overlapping motions, and Zero-Fi targets zero-shot activity recognition, whereas SignalLang learns a unified representation supporting captioning, bidirectional retrieval, and WiQA from segmented CSI.

To evaluate our activity-aware alignment design, we compared SignalLang with a plain BLIP-style baseline using only WTC, WTM, and LM, without auxiliary activity supervision or class-aware negative masking. SignalLang improves seen Wi2TR/T2WiR R@1 from 39.7%/42.4% to 71.5%/89.8%, and multi-reference CIDEr from 114.3 to 129.8. These results, together with the individual component ablations, support the effectiveness of our activity-aware alignment design.

We also retrained HoloLLM under our matched XRF55 protocol, excluding the 10 held-out activities from training. SignalLang achieves higher seen caption LLM-judge accuracy (60.6% vs. 27.7%), multi-reference CIDEr (129.8 vs. 101.8), and seen/unseen WiQA accuracy (80.8%/46.2% vs. 76.7%/39.4%), despite using a substantially smaller language model (137M vs. 7B parameters).

2. Activity-Level Semantics and Instance-Level Evaluation

We acknowledge that our retrieval protocol evaluates activity-level semantics rather than instance-specific matching. XRF55 consists primarily of scripted repetitions of predefined activities, where recordings of the same class often exhibit similar motion patterns and semantically compatible captions. Consequently, exact-pair retrieval may incorrectly penalize valid matches from the same activity. Our current evaluation therefore focuses on activity-level motion semantics rather than recording-specific motion differences.

Regarding the suggested classifier-to-caption baseline, a conventional closed-set classifier maps predicted activity labels to fixed caption templates and cannot directly handle our 10 held-out activity classes without additional zero-shot recognition mechanisms. In contrast, SignalLang generates captions conditioned on WiFi representations rather than simply mapping predicted activity labels to predefined templates.

Our FullCap–SimpleCap comparison (Appendix Table 9) further evaluates the benefit of per-recording caption supervision over a single canonical caption per activity class. FullCap improves caption-based WiQA accuracy from 58.3% to 67.3% on seen activities and from 35.3% to 43.5% on unseen activities, corresponding to improvements of 9.0 and 8.2 percentage points, respectively. These results support the value of per-recording motion-focused language supervision compared with fixed class-level descriptions.

We acknowledge that this comparison does not fully disentangle recording-specific motion information from linguistic diversity, and that our current retrieval evaluation does not establish within-class instance discrimination. We will clarify this distinction in the revised paper.

AuthorsOpenReview ↗

Author Response to Reviewer V2N3 (Continued)

3. WiQA Shortcut Controls

To assess whether WiQA predictions depend on valid CSI inputs rather than linguistic priors, we evaluated question-only (questions and options without CSI), shuffled-CSI (CSI from another recording), and zero-CSI (all-zero input) settings. We also included majority-answer and ground-truth activity-label-only controls. The results are summarized below:

ConditionSeen body partSeen motion attributeUnseen body partUnseen motion attributeAll
Question-only (Qwen3.6, no signal)54.634.838.933.539.7
Majority answer per question67.031.239.828.643.8
Label-only (GT activity → most frequent training answer)94.886.7n/an/a—
SignalLang84.469.561.130.072.7
Shuffled CSI (another recording's signal)52.828.237.828.436.2
Zero CSI67.029.539.830.743.5

Note: All values are accuracies (%). The "All" column includes all four WiQA question types. The label-only diagnostic is provided with ground-truth activity labels rather than inferring them from CSI, and is not defined for unseen activities.

SignalLang achieves substantially higher overall accuracy than the question-only and signal-replacement controls (72.7% vs. 39.7%, 36.2%, and 43.5%), demonstrating that its WiQA predictions depend on valid CSI inputs. In particular, seen body-part and motion-attribute accuracy reach 84.4% and 69.5%, compared with 52.8% and 28.2% under shuffled CSI. On unseen activities, body-part accuracy remains substantially higher (61.1% vs. 37.8%), although unseen motion-attribute accuracy remains limited.

The GT activity-label-only diagnostic achieves 94.8% and 86.7% on seen body-part and motion-attribute questions. Unlike SignalLang, this diagnostic is provided with ground-truth activity labels, serving as an oracle-label reference for activity-level prediction rather than a matched CSI-based baseline. Its high accuracy indicates that many WiQA answers are strongly associated with activity identity. Importantly, SignalLang does not receive activity labels at inference, and its improvements over question-only and shuffled-CSI controls demonstrate that its predictions depend on information extracted from WiFi signals. These findings support CSI-dependent activity-level motion understanding, while recording-specific motion understanding remains to be established.

4. Validity of Motion-Focused Supervision

Our motion-focused supervision aims to retain motion-related information while excluding visual attributes that CSI cannot directly observe, such as clothing color and scene appearance. As shown in Appendix Table 8, the reference-unmatched visual-term rate decreases from 74.3%/77.1% to 3.4%/4.5% on seen/unseen activities, with two video-verified examples in Figure 4. This lexicon-based measure is a proxy rather than a verified factual-error rate. Our additional WiQA controls further show that body-part accuracy substantially exceeds the shuffled-CSI baseline on both seen (84.4% vs. 52.8%) and unseen activities (61.1% vs. 37.8%), supporting CSI-dependent activity-level motion understanding.

The lower LLM classification accuracy under motion-focused supervision can partly be explained by the removal of activity-revealing visual cues. Even for ground-truth captions, the LLM judge achieves 82.6% activity classification accuracy with motion-focused descriptions, compared with approximately 95% for unrestricted captions, which often include informative objects or scene details. Meanwhile, activity-related WiQA accuracy remains similar (91.4% vs. 92.5%), and a length-matched control shows the same trend. These findings suggest that the lower LLM classification accuracy reflects, at least in part, differences in caption content rather than necessarily weaker WiFi representations.

AuthorsOpenReview ↗

Author Response to Reviewer V2N3 (Continued, Part 2)

5. Domain Generalization and Dataset Provenance

We acknowledge that our original repetition-level split does not explicitly hold out subjects or environments, and therefore does not establish generalization to unseen participants or scenes. Cross-subject and cross-environment generalization are well-recognized challenges in WiFi sensing (Zheng et al., 2019; Yang et al., 2023), as CSI measurements depend not only on human motion but also on participant-dependent motion patterns and environment-specific multipath propagation. A comprehensive evaluation of robustness across unseen participants and environments remains an important direction for future work. We will clarify this limitation in the revised paper.

The 28,759 WiFi–text pairs are derived from synchronized XRF55 CSI and videos across four indoor scenes, comprising 16,176 training, 2,697 validation, 8,087 seen-test, and 1,799 unseen-test recordings. For each recording, Qwen3.6-35B generates a motion-focused caption from 16 uniformly sampled video frames and the corresponding activity label, followed by rule-based validation and regeneration. Each caption is paired with its corresponding synchronized CSI recording. Videos are used only for offline annotation, and test annotations are excluded from model training. We will further clarify participant/scene coverage and the recording-pairing procedure to facilitate reproducibility.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.