{"id":"914783ff-98c8-4335-b662-2f9ec08e8d53","arxiv_id":"2505.02052","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A bidirectional text-pressure model with a learned codebook generates synthetic pressure data from activity descriptions and classifies real pressure sequences via LLM-generated text, gaining up to 12.4 macro-F1 points on one benchmark but losing on sleeping-posture data.","lead":"TxP trains two linked models: one that turns text descriptions of movements into synthetic pressure-sensor maps, and one that turns pressure maps into text and activity labels. Adding its synthetic data to real training data improved activity recognition on three of four real datasets, with a best gain of about 12 macro-F1 points.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 12.4% SOTA gain is the best cell from a per-dataset ratio sweep selected on test data, and the three datasets where it holds share the sensor geometry of the synthetic-data simulator; validation-based ratio selection is required.","rationale":"The reader's simulator-fidelity concern is real; the PmatData drop from 0.765 to 0.544 demonstrates it. I place emphasis slightly differently: the more immediately testable threat to the abstract's numerical claim is the evaluation protocol. Tables 3 and 5 present the full ratio sweep, and the headline is the maximum of that sweep. Since no validation-based selection is described, and the best ratio is dataset-dependent (PmatData prefers 100% real), the 'up to 12.4%' number cannot be separated from selection on the test labels. This matters even if PresSim were a perfect simulator. The sensor-geometry issue compounds it: the three positive datasets share the sensing mat geometry that PresSim models, while PmatData and PID4TC show weak or negative transfer. I credit the authors for explicitly listing these limitations (§5.3, §5.4), and the architectural contribution is coherent, so I would not move to reject. The correct status is CONDITIONAL, as the reader stated; I would sharpen the conditions to require validation-based ratio selection, released artifacts, and at least one successful out-of-geometry transfer.","tokens_in":32515,"tokens_out":10472,"duration_ms":113642,"concrete_test":"Split each real dataset into train/validation/test. Select the mix ratio and classifier (baseline 3D-CNN vs Pressure2Text) that maximize macro-F1 on the validation split only, then report macro-F1 on the untouched test split for that configuration. In addition, apply the same validation-selected configuration to PmatData and PID4TC, which cannot be obtained by cropping/resampling the 80×28 SensingTex grid. If the selected ratio is not 50/50 for the headline TMD/MeX cells, or if the out-of-geometry macro-F1 gain over the real-only baseline is within one standard deviation or negative, the claimed SOTA improvement is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline 12.4% macro-F1 improvement is the best cell of a per-dataset real/synthetic ratio sweep. Section 4.5 (Tables 3 and 5) reports ratios from 90/10 through 0/100 and concludes that 50/50 'achieved the highest performance across datasets,' but no validation split or ratio-selection rule is described; PmatData is actually best at 100% real (0.765) and falls to 0.544 at 50/50. Selecting the best ratio on test data turns the reported 'up to' number into a post-hoc maximum, which is not a reproducible configuration rule. The favorable cells are also concentrated on datasets whose sensor geometry is the one PresSim was built to simulate: PresSim, TMD, and MeX all use SensingTex mats (80×28, or crops/resamples of that grid), while the genuinely different PmatData (Vista Medical 32×64) degrades. The paper itself flags the missing lying-down motions in PressLang (§5.3) and the resulting label bias (§5.4), and Table 8 shows only a within-noise gain on the insole dataset PID4TC. The central quantitative claim therefore rests on two unsecured conditions: an honest validation-based choice of the mix ratio, and transfer of PresSim synthetic data to real sensors outside its native geometry.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TxP, a bidirectional Text×Pressure framework for pressure-based HAR. Text2Pressure maps activity descriptions to dynamic pressure sequences via a PressureRQVAE tokenizer and a CLIP-conditioned autoregressive transformer, trained on a new synthetic corpus (PressLang) built by simulating Motion-X SMPL poses with PresSim and re-annotating them with LLaMA. Pressure2Text maps pressure token sequences to activity descriptions through a trainable projection head and a frozen LLaMA 2 13B Chat model, enabling an LLM-grounded classifier. The authors evaluate pressure-map reconstruction, text generation, and downstream HAR on PresSim, TMD, PmatData, MeX, and PID4TC, and report that Text2Pressure augmentation plus Pressure2Text classification improves macro F1 by up to 12.4% over state-of-the-art. The paper includes extensive ablations over codebook size, window size, quantization dropout, real/synthetic ratios, and LLM backbones, and it candidly documents failures on PmatData and on insole data.","tokens_in":32743,"tokens_out":4179,"duration_ms":46051,"significance":"If the central claim held as stated, the contribution would be significant: a text-conditioned synthetic data generator for pressure maps, a language-grounded classifier, and an 81.1K-pair corpus would give the community a new tool for addressing pressure-data scarcity and for interpretable HAR. The architecture is reasonable and the component-level evaluations are thorough, including the useful ablation of LLM backbones and the ethically motivated removal of gender attributes in Pressure2Text (footnote 1). The significance is weakened, however, by two factors that the manuscript itself partly acknowledges: the headline 12.4% is the best cell of a per-dataset ratio sweep with no validation-based selection rule, and the favorable results concentrate on datasets whose sensor geometry is the one PresSim was designed to simulate, while the genuinely different PmatData degrades and the insole dataset gains only within noise. The paper is therefore a promising systems contribution whose broad claims need to be recalibrated and re-evaluated.","major_comments":[{"comment":"The conclusion that the 50/50 real/synthetic ratio 'achieved the highest performance across datasets' is contradicted by the paper's own data: for PmatData in Table 3, macro F1 falls from 0.765 at 100% real to 0.544 at 50/50, and every augmentation ratio including synthetic data is worse than real-only data. The headline 12.4% gain is the best cell of an eight-ratio sweep (TMD, Pressure2Text, 50/50), not a configuration selected by a described validation procedure. Please provide a validation-based rule for choosing the mix ratio, or explicitly present the sweep as exploratory and avoid selecting the maximum on the test sets.","section":"§4.5, Tables 3 and 5"},{"comment":"The transfer claim is load-bearing and currently supported only for datasets whose sensing hardware matches the PresSim simulation target. PresSim, TMD, and MeX use SensingTex mats (80×28, or crops/resamples of that grid), while PmatData (Vista Medical 32×64) degrades under augmentation and PID4TC (insole) improves only from 0.731 to 0.744, within the reported ±0.035 standard deviation. Section 5.3 attributes the PmatData collapse to the absence of lying-down motions in PressLang, which means the corpus is not yet sufficient for the paper's broad 'advancing pressure-based HAR' claim. Please scope the central claim to the simulated mattress geometry or add evidence on additional sensor configurations.","section":"§5.3, Table 8"},{"comment":"The comparison against 'SOTA' is not controlled: the SOTA rows are numbers taken from the original dataset papers with different classifiers and evaluation protocols, while the TxP rows use the authors' 3D CNN baseline or the Pressure2Text classifier. Since the abstract's 12.4% is framed as a gain over state-of-the-art, the comparison should be re-run under a common protocol, or the claim should be limited to gains over the paper's own baseline. The differences in several cells are also within one standard deviation (e.g., PresSim 0.912±0.024 vs PressureTransferNet 0.911±0.015), so significance testing or confidence intervals should accompany the claim.","section":"§4.5, Table 5"},{"comment":"The paper's own limitation statements in §5.4 describe unaddressed LLM bias in re-annotating descriptions and in translating hard labels to activity descriptions, and the proposed mitigation is manual checking rather than a implemented filter. This is an honest disclosure, but it should be reflected in the conclusions: the synthetic corpus is not yet a verified resource for activities outside a narrow set, and the reported gains may partly reflect bias shared between PressLang and the evaluation datasets' label vocabularies.","section":"§4.5, §5.4"}],"minor_comments":[{"comment":"The paragraph beginning 'This section presents the comprehensive approach...' is duplicated verbatim; please remove one instance.","section":"§3, first paragraph"},{"comment":"The appendix states that the residual quantization content 'originates initially from RQ-VAE [20]', but reference [20] is MoMask; the original RQ-VAE citation is [34] (Lee et al.).","section":"Appendix B"},{"comment":"The text refers to 'TDM [54]' in several places while the table and reference list use 'TMD'; please unify the notation.","section":"§4.5 and Table 3"},{"comment":"'PreSim' should be 'PresSim' in the sentence 'it exceeded real-only data on PreSim, TMD dataset, and MeX...'.","section":"§4.5"},{"comment":"In §4.3, 'different matrices' should read 'different metrics'.","section":"§2 and Table 2"},{"comment":"The label 'LRQVAE' in Figure 2 appears to be a typo for 'RQVAE' or 'PressureRQVAE'.","section":"§3.2, Figure 2"},{"comment":"The abstract says '81,100 text-pressure pairs' while the contributions say '81.1K unique motions and 78 million individual pressure frames'; please clarify whether 81.1K counts motions, text-pressure pairs, or both.","section":"Abstract and §1"}],"recommendation":"major_revision","confidential_remarks":"This is a solid systems paper with a clear architecture and unusually candid limitation sections. My main concern for the editor is calibration of the headline claim: the 'up to 12.4%' number is a post-hoc maximum over a test-set ratio sweep, and the positive results are concentrated on datasets whose geometry matches the simulator. I would not reject the paper, but I would require the authors to add a validation-based ratio-selection protocol and to re-state the contribution as scoped to pressure-mattress data of the simulated type unless they add evidence from more diverse sensor geometries. The reliance on the authors' own prior simulator and datasets is not disqualifying but deserves an explicit statement in the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is the first pressure-to-text model I know of that aligns pressure maps with a pretrained LLM, and the PressLang corpus (81k text-pressure pairs synthesized from Motion-X via PresSim) is a real resource for a field that doesn't have many. Second, the headline 'up to 12.4%' macro F1 gain is not a stable result; it's the best cell of an eight-ratio sweep reported on test data, with no validation-based selection rule, and the datasets where it holds use the same SensingTex sensor geometry the simulator was built for. Treat the number as promising rather than established.\n\nWhat's good: the bidirectional loop is a sensible design. PressureRQVAE adapts MoMask's residual quantization to pressure windows cleanly, and the authors evaluate on five datasets, including an insole set. They also honestly document failure: PmatData (sleeping postures) degrades from 0.765 to 0.544 at the 50/50 mix, and they trace it to missing lying-down motions in PressLang and LLM label bias. That's more transparency than most papers in this area. The paper's own limitations section is candid, and the qualitative examples match the claims.\n\nThe soft spots are real. The ratio sweep issue: Tables 3 and 5 scan ratios from 90/10 through 0/100, and the 50/50 conclusion is read off test performance. PmatData is best at 100% real, so the 'best across datasets' claim fails on one of four. Without a held-out validation split, the 'up to' number is a post-hoc maximum, and the reproducibility of the configuration rule is missing. Second, transfer is partly in-distribution: PresSim, TMD, and MeX are all SensingTex mats. The one out-of-distribution sensor (PmatData, Vista Medical) is exactly where the method collapses. The insole result (PID4TC) is within noise. So the 'works on real pressure sensors' claim needs an external sensor win to be convincing. The SOTA comparisons also lean on the authors' own prior systems and lack per-row citations in places.\n\nI think the central idea survives these objections; the magnitude does not. With released code/data, validation-based ratio selection, and one independently collected pressure dataset, this could be a solid IMWUT/UbiComp paper. As is, it's a competent, honest but over-claimed manuscript. I'd send it to peer review with the expectation of a major revision, not desk reject. It deserves referee time for the corpus and the Pressure2Text direction alone.","headline":"Genuinely new pressure-to-text pipeline and a large synthetic corpus, but the headline accuracy gain is a post-hoc best cell and the transfer test is partly in-distribution.","tokens_in":33391,"tokens_out":2006,"would_cite":false,"duration_ms":21035,"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":"The paper shows that a two-way text-pressure model, trained on 81,100 text-pressure pairs, generates pressure maps from text and classifies them via language, lifting activity recognition by up to 12.4% macro F1.","keywords":["human activity recognition","pressure sensors","ground pressure dynamics","text-to-pressure generation","pressure-to-text generation","vector quantization","large language models","synthetic data augmentation"],"falsifier":"Run the same Text2Pressure augmentation on a real mattress dataset whose activity classes all appear in the PressLang motion corpus; if the 50/50-mix macro F1 does not exceed the real-only baseline by roughly the reported margin, or turns negative, the claim that simulated text-to-pressure data transfers to real recordings would be falsified. A cheaper, direct check: generate Text2Pressure maps for the 11 PmatData sleeping classes and measure per-pixel overlap with real PmatData maps—the paper's own numbers already indicate the synthetic maps are poor proxies.","tokens_in":32231,"feed_emoji":"📈","tokens_out":11426,"duration_ms":97011,"temperature":0.7,"pith_summary":"The paper tries to establish that natural language can act as the interchange medium for pressure-sensor data: a model trained on synthetic text-pressure pairs can turn activity descriptions into realistic ground-pressure dynamics, and turn real pressure sequences back into activity descriptions and class labels. If true, this would let researchers generate training data for pressure-based activity recognition from text alone, bypassing slow and costly real sensor collection, and would replace fixed activity labels with open-ended, interpretable descriptions. The headline result is that the two directions combined, with a 50/50 mix of real and synthetic data, raise macro F1 on real mattress datasets by up to 12.4% over prior state-of-the-art classifiers. The paper is explicit that the gain is conditional: on sleeping-posture data, augmentation actually lowers performance because the synthetic corpus lacks lying-down motions.","feed_headline":"Two-way text-pressure model lifts activity recognition by up to 12.4%","feed_subtitle":"Text generation and LLM classification reinforce each other, beating prior models on three of four real datasets","key_machinery":"The load-bearing mechanism is the pressure codebook: PressureRQVAE, a residual vector-quantized autoencoder that slices dynamic pressure maps into two-second windows and quantizes each into a small set of discrete tokens including an end token, so that continuous sensor streams become a finite vocabulary like words. This discrete tokenization is what lets a CLIP text encoder condition an autoregressive transformer to generate pressure sequences from text, and what lets a frozen LLaMA model, via a learned projection head, generate descriptions from pressure. The second pillar is the PressLang corpus itself—81,100 text-pressure pairs simulated by converting 3D SMPL poses into pressure maps with body-shape variations—since both generators are trained end-to-end on this synthetic data. The machinery fails for motion classes absent from the corpus, as the sleeping-posture results show.","core_discovery":"The central claim is that pressure dynamics can be tokenized into a discrete codebook and then aligned with frozen text models in both directions, making pressure a language-compatible modality. TxP's PressureRQVAE compresses variable-length pressure maps into residual-quantized codebook tokens; a CLIP-conditioned autoregressive transformer (Text2Pressure) predicts those tokens from activity descriptions, and a projection head feeding the same token sequence into a frozen 13-billion-parameter LLM (Pressure2Text) generates atomic-motion descriptions that a prompt-engineered classifier maps to activity labels. Trained on PressLang—81,100 motions simulated from 3D body poses with five body-shape variations per motion—the system beats the previous text-to-pressure generator on generation fidelity and, at the best mixing ratio, raises macro F1 by 11.6% through augmentation alone, by 8.2% through grounded classification alone, and by up to 12.4% when both are combined on the TMD daily-activity dataset. The authors present this as an advance for pressure-based HAR, with the caveat that recognition collapses on sleeping-posture data because such motions are absent from the training corpus.","pith_inferences":["Because the synthetic corpus defines the vocabulary of the codebook, the failure on sleeping postures implies that the bottleneck is corpus coverage rather than the quantization or alignment architecture; a testable extension is to add recumbent motion-text pairs to PressLang and watch PmatData F1 recover.","Since generation is tied to one fixed sensor geometry (an 80x28 SensingTex mat), reported gains may be partly geometry-specific; the paper's own adaptation recipe assumes the target array is a crop or resample of the original, so larger or differently shaped mats are an untested edge case.","The same token-plus-LLM recipe could extend to other spatially rich modalities, but the PID4TC insole result warns that the transfer is not automatic because Pressure2Text was trained on mattress data with environment-anchored positions."],"forward_implications":["Text-based augmentation can cut the cost of pressure-HAR dataset collection: with a 50/50 real-to-synthetic mix, the paper reports up to 12.4% macro-F1 gains over prior state-of-the-art on a daily-activity dataset.","The bridge between synthetic and real pressure data holds only for activities covered by the simulation corpus; expanding to new activity families requires adding their motion-text pairs to the training data, not just fine-tuning the classifier.","Pressure2Text turns classification into a language task, so a single model can output free-form atomic-action descriptions rather than a fixed label set, enabling open-vocabulary recognition and human-readable explanations of a pressure sequence.","The best augmentation ratio is roughly balanced (50% real, 50% synthetic); too much synthetic data degrades performance because the model drifts from real sensor statistics.","LLM-based classification is computationally heavier than a 3D-CNN baseline, but the paper reports it can still run at interactive rates on a consumer GPU."],"supporting_citations":[{"why":"Supplies the physics-based simulation that generates every synthetic pressure map in PressLang, so the whole pipeline inherits its fidelity.","marker":"[55]"},{"why":"Provides the 81.1K motion sequences and language annotations that PressLang is built from.","marker":"[37]"},{"why":"Source of the RQ-VAE quantization scheme that PressureRQVAE adapts for pressure dynamics.","marker":"[20]"},{"why":"Provides the frozen CLIP text encoder that conditions Text2Pressure generation.","marker":"[50]"},{"why":"The template for feeding sensor tokens into a language model, which Pressure2Text follows.","marker":"[14]"},{"why":"The closest prior text-to-pressure generator; TxP's reported generation and augmentation gains are measured against it.","marker":"[54]"},{"why":"The state-of-the-art pressure augmentation baseline that TxP must beat for the 12.4% claim.","marker":"[53]"},{"why":"The cross-modal LLM baseline that Pressure2Text is compared with for text-generation quality.","marker":"[77]"}],"fun_headline_variants":["Bidirectional text-pressure model boosts activity recognition by 12.4%","TxP: Turning pressure maps into text improves HAR by up to 12.4%","Pressure-as-language model lifts activity recognition up to 12.4%","Text-to-pressure and back: HAR gains up to 12.4% F1","LLM reads pressure dynamics, boosting activity recognition by 12.4%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pressure maps simulated by PresSim from 3D body poses faithfully reproduce real ground-pressure dynamics for the target activities and sensor layout, so that classifiers trained with text-generated synthetic data transfer to real recordings; the paper itself shows this premise fails for sleeping postures, where augmentation drops macro F1 from 0.765 to 0.544.","fun_headline_variants_meta":{"raw":{"variants":["Bidirectional text-pressure model boosts activity recognition by 12.4%","TxP: Turning pressure maps into text improves HAR by up to 12.4%","Pressure-as-language model lifts activity recognition up to 12.4%","Text-to-pressure and back: HAR gains up to 12.4% F1","LLM reads pressure dynamics, boosting activity recognition by 12.4%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1580,"prompt_tokens":1025,"completion_tokens":555,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":451}},"tokens_in":641,"tokens_out":555,"duration_ms":5194,"temperature":1.0,"reasoning_tokens":451,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:03:19.356162+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same Text2Pressure augmentation on a real mattress dataset whose activity classes all appear in the PressLang motion corpus; if the 50/50-mix macro F1 does not exceed the real-only baseline by roughly the reported margin, or turns negative, the claim that simulated text-to-pressure data transfers to real recordings would be falsified. A cheaper, direct check: generate Text2Pressure maps for the 11 PmatData sleeping classes and measure per-pixel overlap with real PmatData maps—the paper's own numbers already indicate the synthetic maps are poor proxies.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the physics-based simulation that generates every synthetic pressure map in PressLang, so the whole pipeline inherits its fidelity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The closest prior text-to-pressure generator; TxP's reported generation and augmentation gains are measured against it."}],"review_version":1}