PaperlayerView exact revision on arXiv

Source-linked AI summary

CSI-TextBench: A Dataset and Benchmark for Language-Grounded Ambient Sensing Perception

Guozhen Zhu, Yuqian Hu, Sakila S Jayaweera, Wei-Hsiang Wang, Beibei Wang, K. J. Ray Liu

arXiv:2610.34616v1cs.AIcs.ET

TL;DR

CSI records indoor context through widely deployed WiFi, but existing CSI datasets lack natural-language descriptions for language-grounded sensing. CSI-TextBench builds a multitask CSI–text benchmark and evaluates alignment methods; alignment approaches supervised recognition while supporting compositional and open-vocabulary uses.

  • Problem

    CSI datasets provide coarse categorical labels rather than the semantically rich, compositional natural-language descriptions needed to align CSI with language.

    Source: §1 · p. 1

  • Method

    CSI-TextBench unifies three CSI benchmarks and pairs their samples with hierarchically structured, label-derived descriptions through many-to-many assignment.

    Source: §1 · p. 2

  • Results

    CSI–language alignment matches supervised classifiers, transfers to unseen classes with few examples, and supports open-vocabulary prompts; one model achieves 86.9% top-1 accuracy on 140 composite scenes.

    Source: §1 · p. 2, §5.2 · p. 6

  • Takeaways & Limitations

    The released dataset, code, and benchmark provide a foundation for reusable, queryable, language-compatible ambient intelligence on widely available wireless infrastructure.

    Source: §1 · p. 2

    Open-vocabulary results concern unseen phrasing of concepts represented in training, while transfer to unseen classes remains limited on fine-grained tasks.

    Source: §6 · p. 9

Abstract

from arXiv · show

Indoor context deciphered by ambient sensing underpins many applications, from assisted living to smart buildings, yet remains weakly connected to the language-aligned paradigm of modern multimodal systems. Through billions of IoT devices, ambient WiFi Channel State Information (CSI) captures the indoor context of human activity and environmental dynamics passively via ubiquitous, always-on infrastructure. However, unlike vision or audio, CSI lacks naturally paired text like captions or transcripts, limiting its integration with language-driven AI systems. We present CSI-TextBench, a large-scale dataset and benchmark for aligning CSI data with natural language, unifying 3 sensing benchmarks into 285,790 samples across 7 tasks and pairing them with 2,291 hierarchically structured, label-derived descriptions via many-to-many assignment to construct 2.0 million CSI-text training tuples. Benchmarking 5 alignment methods from 4 paradigm families, we show that contrastive CSI-text alignment matches supervised classifiers without class labels, adapts to unseen classes with few samples, and responds to open-vocabulary natural-language prompts. A single unified model supports compositional scene understanding, answering joint queries over activity, proximity, and identity, achieving 86.9% top-1 accuracy on 140 composite scenes, outperforming task-specific classifier ensembles. CSI-TextBench enables ubiquitous ambient sensing as a language-indexed perception layer, enabling indoor contextual awareness in language-driven AI systems.

1 Introduction

Indoor WiFi CSI can passively encode activity, proximity, and identity, but CSI datasets lack the natural-language descriptions needed for language alignment. CSI-TextBench pairs CSI with label-derived text and evaluates language-grounded sensing, including compositional understanding.

1 Introduction

  • WiFi is an always-on, widely deployed sensing modality whose CSI is sensitive to human motion and implicitly encodes activity, proximity, and identity.This makes it a practical basis for passive, device-free ambient intelligence using existing infrastructure.Source: §1 · p. 1
  • Existing WiFi sensing systems often use fixed label spaces and predefined targets, while CSI datasets lack the semantically rich text available for images or speech.This limits how CSI can be queried, reused, and composed as a general-purpose perception interface.Source: §1 · pp. 1–2
  • 86.9% top-1 accuracy over 140 composite scenes shows a single model can answer joint queries over activity, proximity, and identity.The benchmark also tests zero-shot classification, few-shot transfer, and open-vocabulary prompting.Source: §1 · p. 2
  • CSI-TextBench unifies three public benchmarks into 285,790 samples across 7 tasks, with 2,291 descriptions forming approximately 2.0M CSI–text training tuples.Its contributions also include a controlled comparison of five alignment methods and standardized evaluation protocols.Source: §1 · p. 2
  • The benchmark is intended as a foundation for reusable, queryable, language-compatible ambient intelligence on widely available wireless infrastructure.The paper states that the dataset, code, and benchmark are released.Source: §1 · p. 2

2 Related Work

WiFi sensing spans many tasks and datasets, but annotations remain categorical rather than language-grounded. Multimodal alignment offers queryable shared representations, while CSI presents distinct pairing and reuse challenges.

2.1 WiFi sensing

  • Public WiFi benchmarks span diverse devices, environments, and semantic granularities, but provide categorical annotations rather than natural-language descriptions.Prior WiFi sensing includes activity, gesture, fall, localization, proximity, motion-source, and user-identification tasks.Source: §2.1 · p. 2

2.2 Multimodal alignment

  • Language-alignment methods developed for vision, audio, and other modalities use naturally co-occurring text, whereas CSI requires label-derived language and dense many-to-many correspondence.Prior approaches include contrastive, sigmoid, anchor-based, cross-correlation, and masked cross-modal methods.Source: §2.2 · p. 3
  • Related RF-language and other sensing-language studies are often tied to a single task, modality, or application, or do not release reusable sample–text pairs.CSI-TextBench provides an open, multitask WiFi CSI–text benchmark spanning diverse tasks, semantic granularities, and alignment objectives.Source: §2.2 · p. 3

3 The CSI-TextBench Dataset

CSI-TextBench converts heterogeneous public CSI datasets into a unified signal–text resource. It standardizes CSI inputs, generates structured label-derived descriptions, and uses many-to-many pairing to support single-attribute and scene-level alignment.

3 The CSI-TextBench Dataset

  • The benchmark covers 285,790 task-level instances across 7 tasks, with 2,291 descriptions paired through many-to-many assignment to yield approximately 2.0M training tuples.The three multitask CSI-Bench tasks share recordings; the table caption states that the task-level instances correspond to 174,422 unique recordings.Source: §3 · p. 3, Table 2 · p. 4
  • The three source datasets span commercial-device in-the-wild data, controlled indoor activities, and 276 sign-language gesture classes.CSI-Bench supplies five tasks, XRF55 supplies 55 activities, and SignFi supplies the fine-grained gesture task.Source: §3 · p. 3
  • Preprocessing standardizes inputs as amplitude tensors X ∈ R^(C×K×T), with per-sample normalization and subcarrier alignment to a fixed K.The source datasets differ in subcarrier counts, antenna configurations, and sampling rates; C, K, and T denote channels, aligned subcarriers, and temporal length.Source: §3 · p. 4
  • Hand-written templates derive multiple semantically equivalent descriptions from class labels, with Level 1 captions for single attributes and Level 2 captions for combined scene attributes.Descriptions are restricted to information observable from CSI and exclude deployment-specific device and room details.Source: §3 · p. 4
  • Each CSI sample matches 4–9 descriptions, and each description can match many samples, creating ambiguity in positive and negative pairs.This differs from standard one-to-one image–text pairing and makes the benchmark a testbed for alignment under many-to-many correspondence.Source: §3 · p. 4
  • CSI-TextBench is a unified large-scale benchmark for CSI–language alignment across heterogeneous sensing conditions, semantic scopes, and dense many-to-many relationships.For SignFi, the release is limited to preprocessing scripts and CSI–text registry files because of source-dataset restrictions.Source: §3 · pp. 3–4

4 Benchmark Alignment Methods

The benchmark compares five alignment methods under shared CSI–text architecture choices, addressing dense many-to-many correspondence and the absence of naturally paired CSI–text data. A frozen pretrained text encoder keeps learning focused on CSI representations.

4.1 Shared Architecture

  • All methods use a dual encoder with CSI and text projection heads, and both embeddings are L2-normalized before alignment losses are computed.The default shared embedding dimension is 256.Source: §4.1 · pp. 4–5
  • The text encoder is frozen to preserve pretrained semantic structure and isolate cross-modal alignment from language-modeling effects.This setup makes learning occur primarily in the CSI encoder.Source: §4.1 · p. 5

4 Benchmark Alignment Methods

  • Five methods spanning four paradigm families are compared under a unified benchmark to probe alignment assumptions under dense many-to-many correspondence and limited naturally paired CSI–text data.The comparison is designed to isolate alignment approaches for CSI’s distinct signal characteristics.Source: §4 · p. 4, §4.2 · p. 5

4.2 Alignment Objectives

  • The methods cover contrastive, anchor-based, non-contrastive, and reconstructive objectives, including approaches that avoid explicit negative sampling or reconstruct text embeddings from masked CSI.The contrastive variants include InfoNCE and sigmoid losses; the other families include LanguageBind adaptation, Barlow Twins adaptation, and masked cross-modal prediction.Source: §4.2 · p. 5

5 Experiments

The experiments test CSI–text alignment across zero-shot recognition, retrieval, generalization to unseen classes and language, unified multitask sensing, and compositional scene understanding. CLIP InfoNCE performs strongly across these settings, while results also show sensitivity to alignment objectives, text encoders, and description granularity.

5.1 Experimental Setup

  • The experiments compare five alignment methods across seven tasks, using zero-shot prototype matching and few-shot logistic-regression probes on frozen embeddings.The supervised baseline uses the same encoder with a linear head, providing a matched-capacity reference.Source: §5 · p. 5, §5.1 · pp. 5–6

5.2 Layer 1: Alignment Effectiveness

  • CLIP InfoNCE approaches the matched-encoder supervised classifier in zero-shot classification, coming within 0.3 points on three tasks and 2.4 points on 276-class SignFi.Four of five methods exceed 92% on simpler tasks, but sigmoid and asymmetric losses are unstable on SignFi, while Barlow Twins collapses on three tasks.Source: §5.2 · p. 6, Table 3 · p. 6

5.3 Layer 2: Generalization

  • On unseen SignFi classes, zero-shot accuracy is 11.3% (8.5× chance), while few-shot probing reaches 54.0% with one example and 74.6% with ten.Retrieval places the correct unseen class at median rank 12 of 76, with R@10 of 46.1%.Source: §5.3 · pp. 6–7, Table 5 · p. 7
  • Unseen layperson prompts achieve 96.0% accuracy versus 95.9% for training templates, while anchored expert and sensor-grounded prompts reach 96.2% and 77.1%.Raw expert and sensor-grounded prompts score 59.1% and 35.4%, respectively, indicating that performance varies with prompt register and anchoring.Source: §5.3 · p. 7, Table 6 · p. 7

5.4 Human-Written Free-Form Language

  • Template supervision reaches 95.0% on held-out human-written free-form prompts, while broader free-form supervision improves raw expert and sensor-grounded accuracy by 27.9 and 44.8 points.Free-form supervision reduces training-template, layperson, and held-out free-form accuracy by 1.0–2.1 points, trading ordinary-language performance for specialized-language coverage.Source: §5.4 · pp. 7–8, Table 7 · p. 8

5.5 Layer 3: Ambient Scene Perception

  • A single model preserves at least 95% of single-task accuracy on six of seven tasks, including 96.0% on HAR and 96.1% on 276-class SignFi.Proximity falls from 94.6% to 84.5%; the model uses text prototypes rather than task-specific heads or routing.Source: §5.5 · p. 8, Table 8 · p. 8
  • 86.9% top-1 accuracy for composite-scene retrieval exceeds 70.4% joint accuracy from independent per-attribute QA across activity, proximity, and identity.Composite retrieval also reaches 96.9% top-3 accuracy and 0.919 MRR, without task-specific heads or explicit compositional supervision.Source: §5.5 · pp. 8–9, Table 9 · p. 9

5.6 Design Ablations

  • ResNet-18 1D scores 98.7% zero-shot accuracy versus 95.5% for ViT-1D and 94.1% for PatchTST; CLIP text prototypes score 98.7% versus 90.2% with all-MiniLM-L6-v2.Richer compositional templates change single-task HAR and Proximity accuracy by at most 0.3 points, although their main benefit appears in multitask ambient QA.Source: §5.6 · p. 9, Table 10 · p. 9

6 Limitation and Future Work

CSI-TextBench exposes limits in language grounding and transfer: specialized sensor language and fine-grained distinctions remain difficult, and cross-environment and cross-device performance is constrained by WiFi CSI domain sensitivity.

6 Limitation and Future Work

  • Specialized or sensor-grounded language remains difficult for pretrained text encoders, and template-based descriptions may not fully express fine-grained signal distinctions.The authors identify domain-aware text representations and alignment objectives tailored to sensing semantics as future opportunities.Source: §6 · p. 9
  • Open-vocabulary results concern unseen phrasing of concepts represented in training, while transfer to unseen classes remains limited on fine-grained tasks.The authors distinguish novel wording from genuinely unseen class transfer.Source: §6 · p. 9
  • Cross-environment and cross-device performance remains constrained by the domain sensitivity of WiFi CSI.The paper points to model architectures that better bridge signal-level variation with language-level concepts as a future direction.Source: §6 · pp. 9–10

7 Conclusion

The conclusion presents CSI-TextBench as a dataset and benchmark for language-aligned WiFi CSI perception. Across its evaluations, alignment matches supervised classifiers, supports transfer and open-vocabulary queries, and enables compositional understanding.

7 Conclusion

  • 86.9% top-1 accuracy across 140 joint configurations demonstrates compositional scene understanding for activity, proximity, and identity.The conclusion contrasts these capabilities with what the supervised paradigm can structurally provide.Source: §7 · p. 10
  • The benchmark unifies three sensing benchmarks into 285,790 samples across 7 tasks, paired with hierarchically generated descriptions.It supports systematic study of CSI–language alignment.Source: §7 · p. 10
  • Across 5 methods from four paradigm families, language alignment matches supervised classifiers, transfers to unseen classes via few-shot adaptation, and accepts open-vocabulary queries.The dataset, code, and benchmark are released to support research at the intersection of multimodal alignment and ambient sensing.Source: §7 · p. 10

8 Broader Impact

CSI-TextBench frames WiFi CSI as a privacy-preserving, queryable ambient perception layer and supplies data, code, and benchmark protocols for language-grounded sensing. Its captions and evaluations span single-attribute recognition, unseen prompts, and joint scene questions.

8 Broader Impact

  • CSI captures no identifiable visual or audio content, and a shared encoder can answer natural-language questions about who is present, what they are doing, and where.The paper describes assistive applications and smart-home automation, while urging consent mechanisms, on-device processing, and access controls; covert monitoring and non-consensual identification are out of scope.Source: §8 · p. 10

A Preprocessing Details

  • Amplitude-only preprocessing, per-sample normalization, and fixed-size subcarrier standardization produce a shared tensor format for training one CSI encoder across datasets.The pipeline segments recordings into non-overlapping 5-second windows and uses zero-padding or uniform subsampling to standardize subcarrier counts.Source: A Preprocessing Details · p. 14

B Alignment Method Formulations

  • Five alignment objectives compare alternative CSI–text matching assumptions, including contrastive, sigmoid, anchor-based, cross-correlation, and masked cross-modal approaches.SigLIP uses independent sigmoid losses over pairs, which the authors say is better suited to multiple valid matches per sample; masked prediction combines text-embedding reconstruction with InfoNCE.Source: B Alignment Method Formulations · pp. 14–15

C.1 Level 1: Base Captions

  • Label-derived captions provide Level 1 single-attribute descriptions, with manually written paraphrases for each class and many-to-many pairing at the sample level.The text-generation pipeline uses Level 1 captions across tasks and adds Level 2 composite captions for multitask evaluation.Source: C Text Generation Pipeline Details · p. 15, §C.1 · pp. 15–16

C.2 Level 2: Composite Captions

  • Level 2 captions combine activity, proximity, and identity across 140 attribute combinations, with four composite descriptions per scene configuration.They enable scene-level reasoning and help ground non-semantic identity labels alongside interpretable context.Source: §C.2 · p. 16

D.1 Zero-Shot Classification

  • The evaluation suite tests zero-shot recognition, generalization to unseen language, and multi-attribute scene understanding; zero-shot class prototypes average projected Level 1 descriptions and are compared by cosine similarity.The standard zero-shot protocol reports accuracy and weighted F1 on shared test sets.Source: D Evaluation Protocol Details · p. 17, §D.1 · p. 17

D.2 Open-Vocabulary Evaluation

  • Open-vocabulary evaluation replaces training templates with disjoint prompt families and keeps the CSI encoder and projection heads fixed during testing.This protocol tests alignment to new language forms unseen during training.Source: §D.2 · pp. 17–18

D.3 Ambient Scene QA

  • Ambient scene QA tests joint scene representation through per-attribute questions and by selecting among 140 full-scene descriptions, counting a composite answer correct only when all three attributes match.The protocols report per-attribute and joint accuracy, plus top-k accuracy, mean reciprocal rank, and decomposed per-attribute accuracy for composite QA.Source: §D.3 · p. 18

E Text-to-CSI Retrieval

Text-to-CSI retrieval tests whether text prototypes retrieve matching CSI samples, while the human-written language layer examines whether alignment transfers beyond templated descriptions. CLIP InfoNCE leads on most retrieval tasks; free-form supervision improves specialized-language coverage with small reductions on ordinary-language prompts.

E Text-to-CSI Retrieval

  • Text-to-CSI Recall@K ranks CSI samples for each class prototype and averages per-class scores normalized by min(K, nrelevant).With many relevant samples per class, the metric reflects top-ranked class dominance; incorrect items entering the top K can make R@5 or R@10 lower than R@1.Source: E Text-to-CSI Retrieval · pp. 18–19
  • CLIP InfoNCE achieves perfect R@1 on 6 of 7 tasks; SigLIP leads on XRF55 at 92.7%.On SignFi, CLIP InfoNCE reaches 99.3% R@5 and 98.2% R@10; Masked Prediction is the strongest alternative there, at 98.2% R@1 and 93.8% R@5.Source: E Text-to-CSI Retrieval · p. 19, Table 12 · p. 19
  • LanguageBind’s text-to-CSI R@1 falls to 80.0% on HAR, 50.0% on Fall Detection, and 25.0% on Proximity.The authors suggest its asymmetric objective supports matching CSI embeddings to text prototypes without equally tight class-wise CSI clustering.Source: E Text-to-CSI Retrieval · p. 19

F Human-Written Free-Form Language Layer

  • Template-trained CLIP InfoNCE reaches 95.0% on held-out free-form prompts, 0.9 points below its training-template accuracy.Free-form supervision raises raw domain-expert, sensor-grounded, and sensor-grounded ensemble accuracy by 27.9, 44.8, and 7.6 points, while lowering training-template, layperson, and held-out free-form accuracy by 1.4, 2.1, and 1.0 points.Source: F Human-Written Free-Form Language Layer · p. 19
  • Replacing CLIP with MiniLM lowers zero-shot accuracy from 98.7% to 90.2%, alongside the free-form results’ evidence that language coverage and text representation contribute to specialized-prompt performance.The comparison uses identical architectures, splits, and optimization settings for the template-supervised and free-form-supervised models.Source: F Human-Written Free-Form Language Layer · pp. 19–20

G Many-to-Many Retrieval Metrics

The section evaluates retrieval when each CSI sample may correspond to multiple valid captions, distinguishing exact paired-description relevance from activity-level relevance.

G Many-to-Many Retrieval Metrics

  • For each CSI sample, the method ranks the 221-caption HAR gallery and measures Hit@1, average precision, and R-Precision, with mAP averaged over queries.Strict relevance counts the sample’s nine assigned captions; class-level relevance counts captions describing its activity.Source: G Many-to-Many Retrieval Metrics · p. 20, Table 13 · p. 20
  • Under class-level relevance, Hit@1 is 0.990, while mAP is 0.804 and R-Precision is 0.668, leaving headroom in ranking all valid descriptions.Strict relevance tests retrieval of paired descriptions; class-level relevance tests whether all semantically valid activity descriptions rank highly.Source: G Many-to-Many Retrieval Metrics · p. 20

H Held-Out Composition

Held-out-combination tests distinguish generalizing to unseen captions from generalizing to unseen CSI scenes. Cross-environment and cross-device results show domain-shift degradation, while the paper also specifies privacy boundaries for intended use.

H Held-Out Composition

  • The held-out protocol withholds 12 of 49 observed combinations while retaining each constituent attribute value in other training combinations.It evaluates both withheld joint captions with recordings retained and fully withheld scenes, on the same 1,306 test samples.Source: H Held-Out Composition · p. 20
  • The full-data reference scores 58.5% on held-out combinations versus 86.9% across all test samples; the held-out set is harder because six combinations involve walking across distances.The authors note that resolving these combinations requires inferring distance, the weakest attribute in Table 9.Source: H Held-Out Composition · pp. 20–21
  • With the joint caption withheld, the model retains 84.6% and 90.5% of full-data accuracy in open-set and closed-set retrieval; with the entire scene withheld, accuracy is 13.9% and 63.5%.Both settings use 1,306 test samples from 12 held-out combinations; the open-set ranks 140 candidates and closed-set ranks 12.Source: H Held-Out Composition · pp. 20–21, Table 14 · p. 21

I Cross-Environment and Cross-Device Transfer

  • Both aligned and supervised models degrade under environment and device shift; at 20 shots per class, aligned accuracy reaches 45.4% and 55.0%, gains of 13.6 and 15.1 points over zero-shot evaluation.The zero-shot aligned model is 5.0 points below the supervised reference under cross-environment shift and comparable under cross-device shift (39.9% vs. 39.6%).Source: I Cross-Environment and Cross-Device… · p. 21, Table 15 · p. 21

J Provenance, Intended Uses, and Privacy

  • CSI amplitude contains no visual or audio recordings, but can still reveal occupancy and activity; the paper excludes covert monitoring and deployment without access control from intended uses.New collection or deployment requires informed consent and compliance with applicable institutional and legal requirements.Source: J Provenance, Intended Uses, and Privacy · pp. 21–22

K Comparison with Published Source-Dataset Results

The unified CSI-TextBench setup remains competitive with published task-specific results, and a visualization qualitatively shows text prototypes near corresponding CSI clusters on XRF55.

K Comparison with Published Source-Dataset Results

  • On Fall Detection, Motion Source, and User ID, the supervised baseline is within 0.6 points of published results; on XRF55 it matches the source paper’s WiFi accuracy.On SignFi, both supervised and zero-shot models exceed the originally reported 94.8%; the comparison is a reference because training pipelines and evaluation settings may differ.Source: K Comparison with Published… · p. 22, Table 16 · p. 22
  • Zero-shot CLIP InfoNCE stays within 0.1–1.7 points of the supervised baseline across tasks, supporting near-supervised performance with a more flexible inference interface.The largest differences from reported CSI-Bench numbers occur on HAR and Proximity, where the supervised baseline is substantially stronger.Source: K Comparison with Published… · p. 22

L Embedding Visualization

  • On the XRF55 visualization, text prototypes tend to lie within or close to their corresponding CSI clusters across 55 fine-grained action classes.Circles denote CSI test samples, stars denote averaged text prototypes, and matching colors identify the same action class.Source: L Embedding Visualization · p. 22

M Training Details and Hyperparameters

The appendix specifies a shared training protocol and hyperparameters for the alignment experiments, alongside checklist disclosures on reproducibility, evaluation variability, scope, and responsible use.

M Training Details and Hyperparameters

  • All alignment models share the same optimizer, scheduler, projection head, and training protocol unless otherwise noted.The supervised baseline uses the same ResNet-18 1D encoder with a linear classification head.Source: M Training Details and Hyperparameters · pp. 22–23
  • Training uses AdamW, cosine annealing with warmup, batch size 128, and up to 200 epochs with validation-loss early stopping.The appendix also specifies gradient clipping, mixed precision, temperature settings, and a normalized two-layer projection head.Source: M Training Details and Hyperparameters · p. 23

NeurIPS Paper Checklist

  • The authors identify the scope of open-vocabulary evaluation and cross-domain transfer as limitations discussed in the paper.Novel-language generalization is evaluated separately using unseen prompts.Source: NeurIPS Paper Checklist · p. 24
  • The paper reports the architecture, hyperparameters, preprocessing, losses, text generation, evaluation protocols, and released code and data for reproducibility.Data splits follow source-dataset protocols, including CSI-Bench cross-environment and cross-device splits.Source: NeurIPS Paper Checklist · pp. 25–26
  • Reported variability is mean ± standard deviation over 3 seeds for zero-shot results, 10 trials for few-shot results, and 5 trials for target-domain adaptation.The source of variability is stated for each result.Source: NeurIPS Paper Checklist · p. 26
  • The checklist reports the compute setup and duration as one NVIDIA A10G GPU with approximately 15–45 minutes per single-task run and 2–4 hours for multitask training.The reported hardware is an AWS SageMaker ml.g5.xlarge instance with 24 GB GPU memory.Source: NeurIPS Paper Checklist · p. 27
  • The paper discusses assistive applications and privacy advantages alongside risks from activity detection and presence monitoring, with consent and on-device processing as mitigations.The checklist describes the released CSI amplitude data and templated descriptions as posing no direct misuse risks comparable to certain high-risk model releases.Source: NeurIPS Paper Checklist · p. 28
  • The new dataset and benchmark are documented with construction, statistics, preprocessing, evaluation protocols, intended uses, and privacy considerations; the paper uses existing public CSI datasets rather than new human-subject research.The core methodology does not involve LLMs; the descriptions are hand-written or human-written without LLM assistance.Source: NeurIPS Paper Checklist · pp. 29–30

Related papers

Loading 2610.34616v1…