{"id":"81974b46-4e64-48ba-9ac4-3aff3093683d","arxiv_id":"2504.14621","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Adding LLM-enhanced text embeddings as a fixed-weight auxiliary input improves wireless human action recognition and temporal action localization across several public datasets.","lead":"This paper adds frozen text embeddings of action class descriptions to existing wireless sensing models and reports accuracy gains on WiFi, RFID, and mmWave benchmarks. It matters because text labels already exist in most datasets, so the approach promises a near-zero-cost performance boost if the gains are real.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-time text retrieval is unspecified; if ground-truth labels select the text embedding, the reported W+T gains are label leakage.","rationale":"Reader's verdict REJECT with high correctness risk is justified. The paper's abstract and Section 5.4 present specific improvements, and all of them assume that a test-time text feature is available. Section 4.1's JSON retrieval is the only mechanism mentioned, but the key used for retrieval at test time is never defined. This is more serious than missing error bars or prompt selection because it determines whether the experiments measure what the paper claims. The paper does contain useful independent pieces: public datasets, a released code link, and systematic ablation of prompt strategies and encoders. Those do not resolve the protocol gap. Other candidate concerns, such as best-result selection on the test set and absence of error bars, would weaken the strength of the conclusions but only after the retrieval protocol is settled. Therefore no change to the reader's verdict is needed: the central claim remains unsupported until the test-time retrieval procedure is specified and validated.","tokens_in":14997,"tokens_out":4191,"duration_ms":41421,"concrete_test":"Open the released repository (github.com/yangzhenkui/WiTalk) and trace the test-time data loader and inference function. Find where the JSON text embedding is loaded for each test sample and record the lookup key. Then rerun the XRFV2 XRFMamba row and the XRF55 Wi-Fi row under three protocols: (1) retrieval by ground-truth label, (2) retrieval by the wireless-only model's predicted label, and (3) a single generic text embedding shared across all test samples. If protocol (1) reproduces the reported gains, they are leakage; if protocol (2) is required, compare against a two-stage baseline that reuses the wireless-only prediction at the same inference cost; if protocol (3) preserves the gains, the text branch is acting as a fixed bias and the framing should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the paper's central empirical claim is that the text branch can be run at test time without access to the ground-truth action label. The paper never specifies this. Section 4.1 says text labels are encoded once and stored as key-value pairs in JSON files for efficient retrieval; Figure 3 and Section 4.2.3 describe fusing the retrieved text embedding into the sensing model. But for a test sample the label is unknown, so the retrieval key must come from somewhere. The paper does not state whether the key is the ground-truth label (leakage), the model's own predicted label (a two-stage pipeline that is never described), or a single fixed prompt. If it is the ground-truth label, then during inference the model receives information derived from the answer, and the reported gains on XRF55 (3.9%, 2.59%, 0.46%), WiFiTAL (4.98%), and XRFV2 (4.02% to 13.68%) are not evidence for text-guided wireless sensing. If it is the predicted label, the W+T comparison needs a matched two-stage baseline and the reported gains need re-estimation. Because every headline number depends on this unstated protocol, the central claim is unsupported as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WiTalk, a text-enhanced wireless sensing framework that augments CSI, mmWave, and RFID based action recognition and temporal action localization models with text embeddings derived from LLM-generated prompts. The method encodes action labels using hierarchical prompts (label-only, brief description, detailed description), embeds them with CLIP (or other text encoders), refines them with multi-head self-attention, and fuses the text features with wireless signal features at a fixed weight. The authors report accuracy gains on XRF55 (WiFi +3.9%, RFID +2.59%, mmWave +0.46%), average mAP gains on WiFiTAL (+4.98%), and average mAP gains on XRFV2 (+4.02% to +13.68%) across multiple TAL backbones, and include an ablation over text encoders and prompt strategies. The central empirical claim is that adding a text branch improves wireless sensing performance without architectural changes or extra data collection.","tokens_in":15259,"tokens_out":4594,"duration_ms":42446,"significance":"If validated, the idea that semantic text prompts can cheaply boost wireless sensing models is practically attractive, and the paper's scope across three datasets, multiple wireless modalities, and several TAL backbones is appropriate. The authors also provide a code repository and include ablations over prompt strategies and encoders. However, the paper's empirical evidence is currently compromised by the absence of a specified inference-time retrieval protocol, the post-hoc selection of the best configuration on the test set, and the lack of variance estimates. These issues affect every headline result, so the significance of the claimed gains cannot be assessed from the manuscript as written.","major_comments":[{"comment":"The paper never specifies how a text feature is obtained for a test sample without using its ground-truth action label. Section 4.1 states that text labels are encoded once and stored as key-value pairs in JSON files for efficient retrieval, and Section 4.2.3 describes fusing the retrieved text embedding with wireless features, but the retrieval key for a test sample is not defined. If the key is the ground-truth label, the reported gains in Tables 1–3 are inflated by label leakage; if the key is a predicted label, the method is a two-stage pipeline that is not described and would require a matched baseline for the W+T comparison. Since every headline result depends on this protocol, the central claim is unsupported as written.","section":"§4.1, §4.2.3, Figure 3"},{"comment":"The best encoder-prompt configuration (CLIP with TCE) is selected using the same test set on which the final gains are reported. Several alternative configurations perform substantially worse than baseline, e.g., Qwen TCE on XRFV2 drops average mAP from 52.61 to 9.36, and LLaMA TCE drops to 13.74. Without a validation split, a pre-registered selection rule, or multiple seeds, the reported gains are optimistic and not properly controlled.","section":"§5.5, Tables 4–6"},{"comment":"The fusion weight of 0.9 for wireless features and 0.1 for text features is introduced as a fixed choice without justification or sensitivity analysis. This weight directly controls the contribution of the text branch, and the reported improvements could depend heavily on this single hand-picked hyperparameter. The paper should present a sweep over the fusion weight or a principled tuning procedure.","section":"§4.2.3"},{"comment":"No error bars, standard deviations, or multiple seeded runs are reported. For improvements as small as 0.46% on XRF55 mmWave and differences of a few mAP points in Tables 2 and 3, the absence of variance estimates makes it impossible to assess whether the observed gains are statistically meaningful or due to run-to-run variation.","section":"§5.3, §5.4, Tables 1–3"}],"minor_comments":[{"comment":"The phrase 'capturing each subcarrier’s frequency response of each subcarrier' contains a redundant repetition of 'each subcarrier'.","section":"§3, 'Principle of WiFi-Based Human Motion Sensing'"},{"comment":"The entry '84.23 81.9275.04' in the CLIP TDE row appears to be missing a space (likely '84.23 81.92 75.04').","section":"Table 5"},{"comment":"The limitation paragraph correctly notes that improvements are inconsistent across signal types and that the mechanism is unclear; this honesty is appreciated, but the limitations should also explicitly address the need to specify the inference-time retrieval protocol.","section":"§6"},{"comment":"Several references contain incomplete or malformed bibliographic metadata (e.g., reference [8] begins with 'Student Member Ieee'), which should be cleaned up.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper addresses a timely and low-cost idea, and the multi-dataset evaluation is broad. The central problem is that the test-time text retrieval procedure is never specified, so the reported gains may be wholly or partly due to label leakage. The post-hoc selection of the best configuration on the test set further undermines the reliability of the headline numbers. If the authors can supply a valid, non-leaking test-time protocol and re-run the evaluation with proper validation and variance reporting, the work could be resubmitted as a substantially revised manuscript; as it stands, the central empirical claim is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the paper has a likeable premise — take frozen text embeddings from LLM-generated action descriptions, fuse them into wireless sensing backbones with a fixed 0.1 weight, and get free accuracy. The experiments are broad: three datasets, seven backbones, three modalities, and the authors release code. That is real work and the paper is honestly written about its own limitations.\n\nThe new part is the three prompt strategies (label-only, brief context, detailed description) applied across many wireless HAR/TAL models. That's incremental over mmCLIP and TS2ACT, which they cite, but the systematic sweep is useful. I also believe the baseline numbers are probably reproducible since they reuse each original method's code.\n\nNow the soft spot. The paper never says how text features are selected at test time. Section 4.1 says text labels are encoded once and stored as JSON key-value pairs, and Figure 3 shows the text embedding being fused into the sensing model. For a test sample the label is unknown, so the key has to come from somewhere. If it's the ground-truth label, then the W+T results are the model being handed the answer and every reported gain is leakage. If it's the model's own predicted label, that is a two-stage pipeline that is never described and the W+T baseline is not apples-to-apples. Either way, the central comparison as written is unsupported. This is not a minor missing implementation detail; every headline number depends on it.\n\nThere are also smaller issues: no error bars or multiple seeds; the best prompt strategy and encoder (CLIP TCE) are selected after seeing the test results; the fusion weight alpha=0.1 is fixed without sensitivity analysis; and some gains are tiny (mmWave 0.46%). The authors themselves admit they don't understand why the text helps, which is fine, but it makes the post-hoc selection more concerning.\n\nBottom line: the idea is plausible and the experiments are extensive, but until the test-time retrieval protocol is specified and the comparison is made leak-free, I would not trust the quantitative claims. If the authors can clarify that they use a single fixed prompt or a class-agnostic embedding, the paper could be a modest contribution. As written, it needs a major revision or a rejection.\n\nI'd send it to peer review rather than desk-reject, because the flaw is identifiable and fixable in principle, and the experimental footprint is substantial. But I would not cite the current numbers.","headline":"Useful idea, but the test-time text retrieval is never specified, so the headline gains may be label leakage; needs clarification before the results can be trusted.","tokens_in":15733,"tokens_out":2841,"would_cite":false,"duration_ms":25279,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"WiTalk claims that a text branch with LLM-generated hierarchical prompts improves wireless sensing accuracy and temporal action localization with no architectural changes or new data collection.","keywords":["wireless sensing","human action recognition","temporal action localization","text prompts","large language models","multimodal fusion","channel state information"],"falsifier":"Run the full WiTalk pipeline but at inference replace the ground-truth label's text embedding with one derived from the model's own predicted label or from a fixed neutral description; if the W+T gains largely disappear, label leakage explains them.","tokens_in":14831,"feed_emoji":"📡","tokens_out":5331,"duration_ms":42912,"temperature":0.7,"pith_summary":"WiTalk seeks to establish that the action labels already sitting in wireless sensing datasets can be converted into a free accuracy boost. The paper combines large-language-model-generated text prompts—label-only, brief context, and detailed descriptions—with an existing wireless sensing model by adding a lightweight text branch and weighted feature fusion. On XRF55, adding the text branch raises accuracy by 3.9% for Wi-Fi, 2.59% for RFID, and 0.46% for mmWave; on WiFiTAL, average mAP rises 4.98%; on XRFV2, average mAP gains range from 4.02% to 13.68%. If true, this gives a no-new-data, no-architecture-change upgrade path for Wi-Fi, RFID, and mmWave activity recognition and localization.","feed_headline":"A text branch boosts wireless sensing by up to 13.7%","feed_subtitle":"WiTalk fuses LLM-written action descriptions into Wi-Fi, RFID, and mmWave models at zero extra data cost.","key_machinery":"The load-bearing mechanism is the text branch: a contextual template plus LLM-generated hierarchical prompts (label-only, brief context, and detailed description), a text encoder producing fixed embeddings cached in JSON, a multi-head self-attention layer, and weighted fusion of text features with wireless features at a 0.1 text weight. This branch carries the argument because it is the only change between the W and W+T runs; all sensing encoders, losses, and hyperparameters stay fixed.","core_discovery":"We claim that the latent text already present in wireless sensing datasets—the action labels—can be turned into a cheap performance booster. WiTalk enriches each label with a large language model under three prompt strategies of increasing semantic richness, encodes the enriched text with a frozen text encoder, refines it with multi-head self-attention, and fuses it with the wireless feature vector at a text weight of 0.1. Across XRF55, WiFiTAL, and XRFV2, this addition raises accuracy or mAP by the reported margins without changing the sensing model's architecture or collecting any new signal data.","pith_inferences":["If the text feature at inference is chosen using the model's own predicted label rather than the ground-truth class name, the reported gains may shrink; a class-agnostic retrieval test would separate genuine semantic regularisation from label leakage.","The prompt strategy that works best appears to depend on modality: RFID benefits more from detailed descriptions, while Wi-Fi prefers the brief-context prompt, suggesting prompt design should be tuned per signal type.","The same branch could be applied to other RF tasks such as gesture recognition or pose estimation, where text labels are equally available and the fusion cost would be the same."],"forward_implications":["Any existing wireless sensing model can be upgraded with text semantics by adding the text branch, with no architectural changes.","The improvement transfers across modalities (Wi-Fi, RFID, mmWave) and across tasks (human action recognition and temporal action localization), so the effect is not specific to one signal type.","Because labels already exist in every dataset, the method promises a zero-data-collection accuracy gain for deployed systems.","Text encoders trained only on images and text—not on wireless signals—still provide useful semantic guidance for wireless features."],"supporting_citations":[{"why":"Supplies the XRF55 benchmark and the DML baseline whose accuracy rises by 3.9%, 2.59%, and 0.46% across modalities.","marker":"[31]"},{"why":"Supplies the WiFiTAL dataset and the WiFiTAD baseline whose average mAP rises by 4.98%.","marker":"[16]"},{"why":"Supplies the XRFV2 dataset and the XRFMamba backbone used for temporal action localization tests.","marker":"[14]"},{"why":"Provides the CLIP text encoder used to turn enhanced prompts into the text features fused with wireless signals.","marker":"[21]"},{"why":"Inspires the multi-head self-attention refinement of text features, as acknowledged in Section 4.2.2.","marker":"[4]"},{"why":"Motivates using LLM-generated detailed action descriptions as fine-grained prompts.","marker":"[12]"},{"why":"Shows that semantically rich label text can support cross-modal learning for human activity recognition, a precedent for using labels as a free signal.","marker":"[38]"}],"fun_headline_variants":["Zero-cost text prompts boost wireless sensing up to 13.7%","LLM-generated labels improve Wi-Fi, mmWave, RFID sensing","Talk isn't cheap: it boosts wireless sensing accuracy","WiTalk: text prompts lift wireless sensing accuracy for free","Text-insight trick sharpens wireless sensing without new data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"At test time, the text feature for each sample must be obtained without peeking at the sample's ground-truth action label, but the paper does not specify how that retrieval is done.","fun_headline_variants_meta":{"raw":{"variants":["Zero-cost text prompts boost wireless sensing up to 13.7%","LLM-generated labels improve Wi-Fi, mmWave, RFID sensing","Talk isn't cheap: it boosts wireless sensing accuracy","WiTalk: text prompts lift wireless sensing accuracy for free","Text-insight trick sharpens wireless sensing without new data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000775,"raw_usage":{"total_tokens":3420,"prompt_tokens":925,"completion_tokens":2495,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":2410}},"tokens_in":541,"tokens_out":2495,"duration_ms":16948,"temperature":1.0,"reasoning_tokens":2410,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:44:08.761418+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full WiTalk pipeline but at inference replace the ground-truth label's text embedding with one derived from the model's own predicted label or from a fixed neutral description; if the W+T gains largely disappear, label leakage explains them.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the WiFiTAL dataset and the WiFiTAD baseline whose average mAP rises by 4.98%."},{"cited_title":"XRF V2: A Dataset for Action Summarization with Wi-Fi Signals, and IMUs in Phones, Watches, Earbuds, and Glasses","cited_arxiv_id":"2501.19034","evidence_quote":"Supplies the XRFV2 dataset and the XRFMamba backbone used for temporal action localization tests."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inspires the multi-head self-attention refinement of text features, as acknowledged in Section 4.2.2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that semantically rich label text can support cross-modal learning for human activity recognition, a precedent for using labels as a free signal."}],"review_version":1}