Pith. sign in

REVIEW 4 major objections 5 minor 90 references

TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read By grounding language-model reasoning in the sensor signal, TRACE-TS jointly predicts activity labels and explains them, setting the best average accuracy/F1 across seven wearable benchmarks.

desk verdict A serious, well-ablated sensor-language reasoning framework whose one load-bearing weakness is that reasoning fidelity is measured against the very teacher traces used for supervision. read the letter →

arxiv 2608.00200 v1 pith:5RB3AWPP submitted 2026-07-31 cs.AI cs.CLcs.ETcs.LG

Sparsh Rastogi , Tanmay Kumar , Baiyu Chen , Jatin Bedi , Zechen Li , Flora D. Salim This is my paper
classification cs.AIcs.CLcs.ETcs.LG
keywords sensor-languagereasoninghumanactivityrecognitionwearabletimeseriesattributiongroundingDAGtracescross-attentionadaptersLLM-as-judgeexplainableAI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

TRACE-TS tries to establish that wearable-activity reasoning can be both accurate and verifiable: it uses attribution maps from an expert classifier to pin reasoning traces to actual sensor regions, distills those structured DAG traces into a compact language model, and shows this grounded supervision improves both classification and trace quality. The framework trains a frozen LLM with gated cross-attention over sensor memory tokens, so at inference a single pass yields the activity label and a provenance-linked explanation without any teacher model or attribution computation. It also introduces Semantic Node Match (SNM), an LLM-judge metric that separates sound reasoning from hallucinated observations better than standard NLG metrics. The key claim is that grounding and structure are not costs but levers: they raise accuracy and macro-F1 above all evaluated baselines.

What carries the argument

The load-bearing mechanism is the attribution-grounded DAG trace. Integrated Gradients and SHAP are fused by geometric mean into a per-timestep, per-channel salience map; the top-k spatio-temporal regions are serialized into a closed evidence vocabulary; the teacher LLM converts these into a four-level DAG (Observation, Inference, Synthesis, Activity) with based_on provenance edges. This trace serves both as supervision for a compact student LM via gated cross-attention adapters over 8 sensor memory tokens, and as the reference for the proposed SNM metric, which matches nodes by Hungarian assignment and LLM-judged semantic equivalence.

What would settle it

Provide a teacher with deliberately misleading attribution maps (e.g., top regions from a different activity class) and show that the student still produces traces humans rate as well grounded; or show that SNM gives high scores to traces whose cited channels are silent while the text is plausible. The paper's own perturbation study suggests SNM would catch the latter, so the stronger test is whether human raters can distinguish attribution-grounded from misgrounded traces at scale, beyond the single-window sample.

Watch

Extended reading notes

Core claim

The paper's central claim is that reasoning traces for wearable time series can be made trustworthy by construction: attribute first, reason over the attributed regions, and distill the result into a small model. TRACE-TS shows that a teacher LLM, prompted only with serialized IG+SHAP salient regions, produces DAG reasoning traces whose observation nodes cite specific channels and temporal intervals, and that a frozen student LLM with gated cross-attention adapters can learn to reproduce these traces from raw sensor windows alone. The DAG structure—Observation → Inference → Synthesis → Activity with based_on provenance edges—turns out to be supervisory sugar: removing it and using free-form

Load-bearing premise

The reference traces and the SNM judge both come from the same teacher LLM, so if the teacher produces fluent but unsupported reasoning (or the attribution map misses the truly relevant channels), the student inherits that error and SNM cannot catch it—the human grounding check covers only one window per activity class (Section 6).

Editorial extensions

If this is right

  • Wearable models can output both a label and a provenance-linked explanation in a single inference pass, with no attribution or teacher compute.
  • Grounded reasoning supervision is not a drag on classification: removing the DAG structure drops average accuracy by over 20 points.
  • Attribution quality directly controls reasoning quality: random or missing attribution collapses observation grounding (SNM-OF1 from 21.49 to 0).
  • Structure-aware metrics like SNM can localize hallucinations to observation, inference, or synthesis level, something surface metrics cannot.
  • Stronger teachers give better traces and better student accuracy, so trace quality scales with teacher reasoning ability.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same attribution→trace→distill pipeline could transfer to other multivariate time-series domains (clinical monitoring, industrial sensors), where the closed-evidence format would need dataset-specific channel vocabularies.
  • SNM could become a general diagnostic for any structured reasoning format, not just sensor traces, if node types and provenance edges are defined.
  • The ablation pattern suggests that reasoning structure acts as a regularizer: by forcing the model to ground each claim, it prevents the LLM from falling back on language priors—a mechanism that could be tested by probing hidden states during reasoning generation.
  • A testable extension: verify whether the student's attention to sensor tokens aligns with the expert classifier's attribution maps on unseen windows—the paper shows alignment qualitatively, but a quantitative correlation would strengthen the grounding claim.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes TRACE-TS, a two-stage framework for joint activity classification and structured DAG reasoning over wearable sensor windows. In Stage 1, an expert HAR classifier's IG+SHAP attributions are serialized and used to prompt a teacher LLM to produce DAG traces (observation/inference/synthesis/activity nodes with based_on provenance edges). In Stage 2, a frozen student LM is adapted via a projector and gated cross-attention to generate these traces from raw sensor inputs; at inference only the student runs, without attribution or teacher guidance. The paper also introduces SNM, an LLM-as-judge metric comparing predicted and reference traces at observation, inference, and synthesis levels. Experiments on seven HAR benchmarks report the best average accuracy and macro-F1 among evaluated methods (84.43%/81.24%), with extensive ablations over attribution source, reasoning format, teacher quality, sensor encoder, token count, adapter rank, and SNM perturbation behavior.

Significance. If the grounding claim were verified, this would be a substantial contribution: it addresses a real limitation of existing LLM-based explanations for time-series data, provides traceable structured reasoning with explicit provenance, and SNM is a sensible diagnostic idea. The paper's strengths include a freely released implementation, coverage of seven benchmarks, controlled ablations, a four-judge robustness analysis for SNM, and a perturbation study showing that SNM is sensitive to certain kinds of misgrounding. However, the central claim that the generated reasoning is "attribution-grounded" is currently underdetermined: the supervision and the evaluation share the same teacher-generated references, the teacher prompt supplies the predicted activity label, the human grounding audit is very small, and classification results are single runs. These issues are partly acknowledged in Section 6 and Appendix D.1, but they need to be addressed experimentally before the paper's main claim can be accepted.

major comments (4)
  1. [§3.1, Fig. 8, Eq. (3), §4.1, §6] The supervision pipeline and the SNM evaluation use the same teacher references. Figure 8 shows that the teacher prompt includes 'Predicted Activity: <ACTIVITY_LABEL>' and asks the model to generate a trace 'for <ACTIVITY_LABEL>'; Eq. (3) trains the student to imitate these label-conditioned traces; and SNM (Sec. 4.1) scores predicted traces against the same teacher-generated references. A student can therefore achieve high SNM by imitating teacher vocabulary and structure, and SNM cannot detect cases where both teacher and student fabricate an observation. Table 8's 'Misgrounded' perturbation demonstrates SNM sensitivity at the inference and synthesis levels (SNM-IF1 −24.35, SNM-SA −72.01), but SNM-OF1 actually increases (+1.18), so observation-level grounding is not independently verified. The paper acknowledges this in Section 6 ('not an absolute verification of reasoning faithfulness
  2. [Appendix D.1, Table 1] All classification results are from a single training run per backbone×dataset combination (Appendix D.1). The central recognition claim rests on average accuracy/F1, but several per-dataset gaps are small (e.g., UCI-HAR 96.67 vs. 97.50 for Attend & Discriminate; USC-HAD 72.35 vs. 73.09 for MantisV2; MHEALTH 98.17 vs. 98.10 for SensorLLM), and OPPORTUNITY is substantially below the specialist best (77.13 vs. 84.33). Without multiple seeds, error bars, or paired significance tests, the claim of 'best average' performance may be within run-to-run variation. Please report at least three seeds with mean±std, or paired tests on the comparisons that matter for the headline claim. The current SNM std values in Table 3 are across judges, not training runs, and therefore do not address this.
  3. [§4.3, Table 4] The human grounding evaluation samples one window per activity class per dataset, with five annotators and no inter-annotator agreement reported. It rates only TRACE-TS traces and does not include baseline traces for comparison. This is too small and too selective to certify grounding across seven benchmarks. The blind pairwise comparison in Table 19 (attribution-grounded vs. no-attribution, 250 judgments) is valuable, but it measures preference between supervision conditions, not whether the cited observations correspond to actual signal events. Please expand the human audit (or a signal-statistics audit) and report IAA; otherwise, claims about 'grounded reasoning' should be softened to expert preference for attribution-conditioned traces.
  4. [§4.1, Appendix H] SNM is a new metric and the paper's main diagnostic instrument, but its validity is established only indirectly: four judges give stable rankings and the perturbation study shows sensitivity to selected rewiring. No correlation with human judgments of trace quality is reported, and the perturbation study does not cover systematic teacher hallucinations (e.g., fluent observations that cite the correct channel but describe a pattern absent from the signal). Given that SNM is used both as a headline contribution and as evidence for the attribution-grounded claim, a direct validation against human annotations or against an independent signal-level audit is needed.
minor comments (5)
  1. [Figure 8] The reproduced prompt has a numbering error: the rule list skips index '11'. The caption notes this, but the authors should either renumber or fix the source prompt in the released code.
  2. [Figure 5(a)] The caption 'Early (ℓ ≤ 1)' is confusing; it presumably means early layers rather than only layer 1. Please clarify the layer grouping.
  3. [§4.4, Table 5] The text says that among faithful attribution sources, SNM differences fall within run-to-run variation, but the paper's reporting convention is single-run. Either add multi-run evidence or rephrase this as exploratory observation.
  4. [Appendix D.2 / Figure 8] The threshold and serialization choices (e.g., 'global p90', 'top-k regions', temporal vocabulary) are described as hyperparameters but their sensitivity is not systematically studied. At minimum, state whether the reported results are robust to reasonable changes in these choices.
  5. [§4.3 / Table 4] The human evaluation aggregates ratings over a very small number of samples; the per-dataset standard deviations are large. It would help to state the total number of rated traces explicitly in the main text and to report the sample size per dataset.

Circularity Check

1 steps flagged · score 4.0 of 10

Reasoning-fidelity evaluation is partially circular: SNM scores predicted traces against the same teacher-generated traces used as training targets, so SNM measures imitation fidelity rather than independent sensor grounding.

  1. fitted input called prediction [Section 3.2 (Eq. 3); Section 4.1 (Eq. 4); Section 6]
    "SNM evaluates a generated reasoning trace against a reference DAG trace... we formulate node matching as an assignment problem between reference nodes O∗ and predicted nodes. The student is trained to generate the teacher trace with a causal language modeling objective... L = −Σ log P(g_m | p, g_<m, Z). Similarly, SNM evaluates generated traces against teacher-generated reference traces, so it should be viewed as a structure-aware diagnostic metric rather than an absolute verification of reasoning faithfulness."

    The reference nodes used by SNM are the same teacher traces g that serve as supervision targets in Eq. 3. The student is optimized to maximize the likelihood of those exact traces, and SNM then scores the student by matching to those same traces. High SNM is therefore, by construction, high agreement with the training target; it does not independently verify that observations are grounded in the sensor signal. If the teacher and student share a hallucinated observation, SNM cannot detect it because the reference contains the same fabrication. The paper's own limitation section concedes this. Classification results are unaffected because they are evaluated against held-out labels, so the circularity is partial.

full rationale

The classification contribution is self-contained: TRACE-TS accuracy/F1 are measured against held-out activity labels on seven external benchmarks, and the ablations (attribution source, output format, encoder/classifier) are controlled comparisons. No load-bearing self-citation chain or imported uniqueness theorem is present. The circular element is confined to the reasoning-fidelity evaluation. SNM's reference traces are the same teacher-generated DAGs used as the Eq. 3 supervision target, so the SNM scores in Table 3 and the perturbation study establish sensitivity and imitation fidelity, not independent sensor grounding. The paper explicitly acknowledges this in Section 6. The blind pairwise human evaluation and the no-attribution ablation provide some non-circular support for the attribution-grounding claim, and the teacher prompt's inclusion of the predicted activity label (Appendix M) further weakens the inference from high SNM to signal-grounded reasoning. On balance, this is partial circularity in the reasoning evaluation loop, not a fully circular derivation; score 4.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

TRACE-TS introduces no new physical entity. Its free parameters are mostly design choices in the evidence-extraction and cross-attention pipeline; the key hidden assumption is that teacher-generated traces are faithful enough to serve as both supervision and evaluation reference. The paper acknowledges the main risk in Section 6 but cannot fully discharge it without independent verification of trace grounding beyond a small human sample.

free parameters (4)
  • Saliency threshold (global p90) = per-sample global p90 threshold
    Attribution regions are selected when mean importance exceeds a data-dependent p90 threshold (teacher prompt in Appendix M). The threshold is chosen by the method, not derived.
  • Top-k salient regions = up to 10 per sample
    The evidence records passed to the teacher are the top-k high-attribution regions; k is a hand-set design choice (Appendix M prompt: 'up to 10').
  • Temporal phase vocabulary and confidence tiers = 6 temporal labels; 3 confidence tiers
    Continuous timesteps and attribution strengths are quantized into a closed vocabulary chosen by the authors, affecting both teacher supervision and SNM reference structure.
  • Sensor memory tokens (N=8) and adapter rank (r=128) = N=8, r=128
    Architectural hyperparameters selected for all experiments; ablation shows performance varies with them, indicating they are tuned rather than forced by theory.
assumptions (5)
  • domain assumption IG and SHAP attribution maps, fused geometrically, identify sensor regions that actually support the true activity.
    Section 3.1 uses attribution from an expert classifier to define salient evidence; if attribution is noisy or misaligned with ground-truth causes, the teacher traces are misgrounded. The paper's own ablations show degraded attribution hurts, but do not establish that fused attribution is objectively correct.
  • domain assumption The teacher LLM follows the DAG schema and does not invent sensor patterns beyond the provided evidence.
    Appendix M attempts to enforce this with prompt rules, but unfaithful generation remains possible; the authors acknowledge in Section 6 that teacher trace errors are inherited by the student.
  • domain assumption Teacher-generated traces are valid reference standards for reasoning evaluation.
    SNM compares student traces to teacher traces; the paper's Section 6 states SNM is not an absolute verification of faithfulness, so the evaluation inherits teacher quality.
  • domain assumption Frozen pretrained MantisV2 encoder and Gemma/Qwen/LLaMA backbones can support sensor-to-language alignment through small adapters.
    The whole cross-attention design assumes the frozen language model can absorb sensor information via gated adapters without catastrophic forgetting; supported empirically per dataset, but not proven in general.
  • domain assumption The public benchmark splits and labels used are trustworthy for measuring human activity recognition.
    The results rely on standard HAR datasets and splits (Appendix A); any label noise or protocol inconsistency propagates into all accuracy/F1 numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding." pith.science (2026). https://pith.science/paper/5RB3AWPP

@misc{pith2026260800200,
  author       = {Pith},
  title        = {Pith review of: TRACE-TS: Attribution-Grounded and Traceable Sensor-Language Reasoning for Human Activity Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5RB3AWPP}},
  note         = {Machine review of arXiv:2608.00200}
}
read the original abstract

Wearable sensors capture fine-grained motion patterns that support rich behavioral understanding, yet most existing methods reduce these signals to activity labels. Recent LM-based approaches generate natural-language explanations for sensor data, but their reasoning is weakly grounded in the underlying signal, leading to fluent yet unverifiable explanations. We introduce TRACE-TS (Traceable Reasoning with Attribution-Grounded Evidence), a framework for structured and signal-grounded reasoning over wearable time series. TRACE-TS uses attribution from an expert classifier to identify salient spatio-temporal sensor regions, uses them to construct DAG reasoning traces with explicit evidence provenance, and trains a compact language model to generate these traces through gated cross-attention over sensor memory tokens. At inference, the adapted model jointly outputs the activity prediction and its reasoning trace, without requiring attribution computation or teacher guidance. We introduce Semantic Node Match(SNM), an LLM-as-judge metric that diagnoses reasoning fidelity at the observation, inference, and synthesis levels, localizing hallucinated observations and broken evidence chains missed by standard NLG metrics. Across seven wearable benchmarks, TRACE-TS achieves the best average accuracy and F1 among all evaluated methods (84.43%/81.24%), and outperforms the best LLM-based baseline by 17.96% in F1. Our code is available at https://github.com/SparshRastogi/TRACE-TS.

Figures

Figures reproduced from arXiv: 2608.00200 by the authors.

Figure 1
Figure 1. Representative Method Families for human activity [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An illustrative representation of the overall methodology of the proposed TRACE-TS framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative reasoning examples and comparisons. We use green to highlight correct parts, red for mistakes. Video is [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Ablation analysis of (a) sensor token count and (b) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: (a) Layer-wise adapter gating and (b) sensor-token [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Per-class mean attribution per sensor channel for [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Fused IG+SHAP attribution heatmaps for two representative activity classes. Rows are sensor channels and columns [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Full, verbatim teacher reasoning-generation prompt for UCI-HAR (Qwen3.5-122B-A10B); the output vocabulary [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

90 extracted references · 2 canonical work pages

  1. [1]

    Miller, Saba Emrani, Ud- hyakumar Nallasamy, and Ian Shapiro

    Salar Abbaspourazad, Oussama Elachqar, Andrew C. Miller, Saba Emrani, Ud- hyakumar Nallasamy, and Ian Shapiro. 2024. Large-scale Training of Foundation Models for Wearable Biosignals. InICLR. https://arxiv.org/abs/2312.05409

    arXiv 2024

  2. [2]

    Sherif El Abd, Vaibhav Aggarwal, Robin Algayres, Alek Andreev, Olivier Bachem, Ian Ballantyne, Brick, et al. 2026. Gemma 4 Technical Report. doi:10.48550/ARXIV. 2607.02770

    arXiv doi:10.48550/arxiv.2607.02770 2026

  3. [3]

    Alireza Abedin, Mahsa Ehsanpour, Qinfeng Shi, Hamid Rezatofighi, and Damith C Ranasinghe. 2021. Attend and discriminate: Beyond the state-of-the-art for human activity recognition using wearable sensors.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies5, 1 (2021), 1–22

    2021

  4. [4]

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al

  5. [5]

    Yair A Andrade-Ambriz, Sergio Ledesma, Mario-Alberto Ibarra-Manzano, Mar- vella I Oros-Flores, and Dora-Luz Almanza-Ojeda. 2022. Human activity recogni- tion using temporal convolutional neural network architecture.Expert Systems with Applications191 (2022), 116287

    2022

  6. [6]

    Davide Anguita, Alessandro Ghio, Luca Oneto, Xavier Parra, and Jorge Luis Reyes-Ortiz. 2013. A public domain dataset for human activity recognition using smartphones. InProceedings of the European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN). 437–442

    2013

  7. [7]

    Maddix, Pablo Guer- ron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael Bohlke-Schneider

    Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guer- ron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael...

  8. [8]

    Mizuki Arai, Tatsuya Ishigaki, Masayuki Kawarada, Yusuke Miyao, Hiroya Taka- mura, and Ichiro Kobayashi. 2025. Evaluating LLMs’ Ability to Understand Numerical Time Series for Text Generation. InProceedings of the 18th Inter- national Natural Language Generation Conference, Lucie Flek, Shashi Narayan, Lê Hong Phuong, and Jiahuan Pei (Eds.). Association fo...

    2025

Show all 90 references
  1. [9]

    Sheikh Asif Imran Shouborno, Mohammad Nur Hossain Khan, Subrata Biswas, and Bashima Islam. 2026. LLaSA: A Sensor-Aware LLM for Natural Language Reasoning of Human Activity from IMU Data. InCompanion of the 2025 ACM International Joint Conference on Pervasive and Ubiquitous Com...

    2026

  2. [10]

    Oresti Baños, Miguel Damas, Héctor Pomares, Ignacio Rojas, Mate Attila Toth, and Oliver Amft. 2014. mHealthDroid: A novel framework for agile development of mobile health applications. InInternational Workshop on Ambient Assisted Living and Home Care. Springer, 91–98

    2014

  3. [11]

    Sizhen Bian, Mengxi Liu, Siyu Yuan, Lala Shakti Swarup Ray, Bo Zhou, Bin Guo, Zhiwen Yu, Thomas Ploetz, Paul Lukowicz, and Vitor Fortes Rey. 2026. Founda- tion Models Defining A New Era In Sensor-based Human Activity Recognition: A Survey And Outlook.arXiv preprint arXiv:2604....

    2026 arXiv

  4. [12]

    Raaj Kishore Biswas, Matthew N Ahmadi, Adrian Bauman, Karen Milton, Nicholas A Koemel, and Emmanuel Stamatakis. 2025. Wearable device-based health equivalence of different physical activity intensities against mortality, cardiometabolic disease, and cancer.Nature Communication...

    2025

  5. [13]

    Yize Cai, Baoshen Guo, Flora Salim, and Zhiqing Hong. 2025. Towards gener- alizable human activity recognition: A survey.arXiv preprint arXiv:2508.12213 (2025)

    2025 arXiv

  6. [14]

    Shing Chan, Hang Yuan, Catherine Tong, Aiden Acquah, Abram Schonfeldt, Jonathan Gershuny, and Aiden Doherty. 2024. CAPTURE-24: A large dataset of wrist-worn activity tracker data collected in the wild for human activity recognition. arXiv:2402.19229 [cs.LG]

    2024 arXiv

  7. [15]

    Youngjae Chang, Akhil Mathur, Anton Isopoussu, Junehwa Song, and Fahim Kawsar. 2020. A systematic study of unsupervised domain adaptation for ro- bust human-activity recognition.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies4, 1 (2020), 1–30

    2020

  8. [16]

    Baiyu Chen, Zechen Li, Wilson Wongso, Lihuan Li, Xiachong Lin, Hao Xue, Benjamin Tag, and Flora Salim. 2026. AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild.arXiv preprint arXiv:2605.22715(2026). https://arxiv.org/abs/2605.22715

    2026 arXiv

  9. [17]

    Baiyu Chen, Wilson Wongso, Zechen Li, Yonchanok Khaokaew, Hao Xue, and Flora Salim. 2026. Comodo: Cross-modal video-to-imu distillation for efficient egocentric human activity recognition.Proc. ACM Interact. Mob. Wearable Ubiq- uitous Technol.(2026)

    2026

  10. [18]

    Linyao Chen, Qinlao Zhao, Zechen Li, Mingming Li, Likun Ni, Jinyu Chen, Yuhao Yao, Xuan Song, Noboru Koshizuka, and Hiroki Kobayashi. 2026. Towards Efficient and Evidence-grounded Mobility Prediction with LLM-Driven Agent. arXiv preprint arXiv:2606.05130(2026). https://arxiv.o...

    2026 arXiv

  11. [19]

    Mathias Ciliberto, Vitor Fortes Rey, Alberto Calatroni, Paul Lukowicz, and Daniel Roggen. 2021. Opportunity++: A multimodal dataset for video- and wearable, ob- ject and ambient sensors-based human activity recognition.Frontiers in Computer Science3 (2021), 792065

    2021

  12. [20]

    Shohreh Deldari, Dimitris Spathis, Mohammad Malekzadeh, Fahim Kawsar, Flora D Salim, and Akhil Mathur. 2024. Crossl: Cross-modal self-supervised learning for time-series through latent masking. InProceedings of the 17th ACM international conference on web search and data minin...

    2024

  13. [21]

    Shohreh Deldari, Hao Xue, Aaqib Saeed, Daniel V Smith, and Flora D Salim. 2022. Cocoa: Cross modality contrastive learning for sensor data.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies6, 3 (2022), 1–28

    2022

  14. [22]

    Sourish Gunesh Dhekane and Thomas Ploetz. 2025. Transfer learning in sensor- based human activity recognition: A survey.Comput. Surveys57, 8 (2025), 1–39

    2025

  15. [23]

    Iveta Dirgová Luptáková, Martin Kubovčík, and Jiří Pospíchal. 2022. Wearable sensor-based human activity recognition with transformer model.Sensors22, 5 (2022), 1911

    2022

  16. [24]

    Vasilii Feofanov, Songkang Wen, Jianfeng Zhang, Lujia Pan, and Ievgen Redko

  17. [25]

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. 2024. MOMENT: A family of open time-series foundation models. InInternational conference on machine learning

    2024

  18. [26]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schel- ten, Alex Vaughan, et al . 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783(2024)

    2024 arXiv

  19. [27]

    Yu Guan and Thomas Plötz. 2017. Ensembles of deep lstm learners for activity recognition using wearables.Proceedings of the ACM on interactive, mobile, wearable and ubiquitous technologies1, 2 (2017), 1–28

    2017

  20. [28]

    Harish Haresamudram, Apoorva Beedu, Varun Agrawal, Patrick L Grady, Irfan Essa, Judy Hoffman, and Thomas Plötz. 2020. Masked reconstruction based self-supervision for human activity recognition. InProceedings of the 2020 ACM International Symposium on Wearable Computers. 45–49

    2020

  21. [29]

    Zhiqing Hong, Zelong Li, Shuxin Zhong, Wenjun Lyu, Haotian Wang, Yi Ding, Tian He, and Desheng Zhang. 2024. Crosshar: Generalizing cross-dataset human activity recognition via hierarchical self-supervised pretraining.Proceedings of the ACM on Interactive, Mobile, Wearable and ...

    2024

  22. [30]

    Hassan Ismail Fawaz, Benjamin Lucas, Germain Forestier, Charlotte Pelletier, Daniel F Schmidt, Jonathan Weber, Geoffrey I Webb, Lhassane Idoumghar, Pierre- Alain Muller, and François Petitjean. 2020. Inceptiontime: Finding alexnet for time series classification.Data mining and...

    2020

  23. [31]

    Sijie Ji, Xinzhe Zheng, and Chenshu Wu. 2024. Hargpt: Are llms zero-shot human activity recognizers?. In2024 IEEE International Workshop on Foundation Models for Cyber-Physical Systems & Internet of Things (FMSys). IEEE, 38–43

    2024

  24. [32]

    Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen

    Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y. Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen. 2024. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. arXiv:2310.01728 [cs.LG] https://arxiv.org/abs/2310.01728

    2024 arXiv

  25. [33]

    Yaxuan Kong, Yiyuan Yang, Yoontae Hwang, Wenjie Du, Stefan Zohren, Zhangyang Wang, Ming Jin, and Qingsong Wen. 2025. Time-MQA: Time Series Multi-Task Question Answering with Context Enhancement. InProceedings of the 63rd Annual Meeting of the Association for Computational Ling...

    2025 doi

  26. [34]

    Ross Koval, Nicholas Andrews, and Xifeng Yan. 2025. Multimodal Language Models for Financial Forecasting from Interleaved Sequences of Text and Time Series. InProceedings of the 14th International Joint Conference on Natural Lan- guage Processing and the 4th Conference of the ...

    2025

  27. [35]

    Harold W Kuhn. 1955. The Hungarian method for the assignment problem. Naval research logistics quarterly2, 1-2 (1955), 83–97. 9

    1955

  28. [36]

    Patrick Langer, Thomas Kaar, Max Rosenblattl, Maxwell A Xu, Winnie Chow, Martin Maritsch, Robert Jakob, Ning Wang, Juncheng Liu, Aradhana Verma, et al

  29. [37]

    Zikang Leng, Hyeokhyen Kwon, and Thomas Plötz. 2023. Generating virtual on-body accelerometer data from virtual textual descriptions for human activity recognition. InProceedings of the 2023 ACM International Symposium on Wearable Computers. 39–43

    2023

  30. [38]

    Zechen Li, Baiyu Chen, Hao Xue, and Flora D. Salim. 2026. ZARA: Training- Free Motion Time-Series Reasoning via Evidence-Grounded LLM Agents. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Maria Liakata, Vivi...

    2026 doi

  31. [39]

    Zechen Li, Shohreh Deldari, Linyao Chen, Hao Xue, and Flora D Salim. 2025. Sensorllm: Aligning large language models with motion sensors for human activity recognition. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 354–379

    2025

  32. [40]

    Opentslm: Time-series language models for reasoning over multivariate medical text-and time-series data.arXiv preprint arXiv:2510.02410(2025)

    2025

  33. [41]

    Chenxi Liu, Shaowen Zhou, Qianxiong Xu, Hao Miao, Cheng Long, Ziyue Li, and Rui Zhao. 2025. Towards Cross-Modality Modeling for Time Series Analytics: A Survey in the LLM Era. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence (IJCAI-2...

    2025 doi

  34. [42]

    Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. 2022. Non-stationary transformers: Exploring the stationarity in time series forecasting.Advances in neural information processing systems35 (2022), 9881–9893

    2022

  35. [43]

    Scott M Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions.Advances in neural information processing systems30 (2017)

    2017

  36. [44]

    Lee, Yuwei Zhang, Maxwell A

    Zechen Li, Keerthana Natarajan, Weizhi Zhang, Menglian Zhou, Simon A. Lee, Yuwei Zhang, Maxwell A. Xu, Zeinab Esmaeilpour, Flora D. Salim, Mark Malhotra, Lindsey Sunden, Shwetak Patel, Yuzhe Yang, and Ahmed A. Metwally. 2026. GlucoFM: A Dual-Stream Foundation Model for Continu...

    2026 arXiv

  37. [45]

    Fink, Sascha Feldhorst, and Michael Ten Hompel

    Fernando Moya Rueda, René Grzeszick, Gernot A. Fink, Sascha Feldhorst, and Michael Ten Hompel. 2018. Convolutional Neural Networks for Human Activity Recognition Using Body-Worn Sensors.Informatics5, 2 (2018). doi:10.3390/ informatics5020026

    2018

  38. [46]

    Metwally

    Hada Melino Muhammad, Zechen Li, Flora Salim, and Ahmed A. Metwally. 2026. CGM-JEPA: Learning Consistent Continuous Glucose Monitor Representations via Predictive Self-Supervised Pretraining. arXiv:2605.00933 [cs.LG] https: //arxiv.org/abs/2605.00933

    2026 arXiv

  39. [47]

    Vishvak S Murahari and Thomas Plötz. 2018. On attention models for human activity recognition. InProceedings of the 2018 ACM international symposium on wearable computers. 100–103

    2018

  40. [48]

    Shenghuan Miao and Ling Chen. 2026. Wonderwall: A Virtual-to-Real Founda- tion Model for IMU-based HAR.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies10, 1 (2026), 1–31

    2026

  41. [49]

    OpenAI. 2025. gpt-oss-120b & gpt-oss-20b Model Card.arXiv preprint arXiv:2508.10925(2025)

    2025 arXiv

  42. [50]

    Francisco Javier Ordóñez and Daniel Roggen. 2016. Deep convolutional and lstm recurrent neural networks for multimodal wearable activity recognition.Sensors 16, 1 (2016), 115

    2016

  43. [51]

    Theresa Patten, Edward A Preble, Hiral Master, Jennifer Adjemian, Andrea Ramirez, James McClain, and Amy Rose Price. 2026. The All of Us Research Program’s wearables dataset.Nature Medicine(2026), 1–9

    2026

  44. [52]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In International Conference on Learning Representations

    2023

  45. [53]

    Jennifer M Radin, Nathan E Wineinger, Eric J Topol, and Steven R Steinhubl. 2020. Harnessing wearable device data to improve state-level real-time surveillance of influenza-like illness in the USA: a population-based study.The Lancet Digital Health2, 2 (2020), e85–e93

    2020

  46. [54]

    Attila Reiss and Didier Stricker. 2012. Introducing a new benchmarked dataset for activity monitoring. In2012 16th International Symposium on Wearable Computers. IEEE, 108–109

    2012

  47. [55]

    Daniel Roggen, Alberto Calatroni, Mirco Rossi, Thomas Holleczek, Kilian Förster, Gerhard Tröster, Paul Lukowicz, David Bannach, Gerald Pirkl, Alois Ferscha, et al. 2010. Collecting complex activity datasets in highly rich networked sensor environments. In2010 Seventh Internati...

    2010

  48. [56]

    Qwen Team. 2026. Qwen3.5: Towards Native Multimodal Agents. https://qwen. ai/blog?id=qwen3.5

    2026

  49. [57]

    Muhammad Shoaib, Stephan Bosch, Ozlem Durmaz Incel, Hans Scholten, and Paul J M Havinga. 2014. Fusion of smartphone motion sensors for physical activity recognition.Sensors14, 6 (2014), 10146–10176

    2014

  50. [58]

    Preetham Sivalingam, Murari Mandal, Saurabh Deshpande, and Dhruv Ku- mar. 2026. LLM-as-a-Judge for Time Series Explanations.arXiv preprint arXiv:2604.02118(2026)

    2026

  51. [59]

    Elena Smets, Emmanuel Rios Velazquez, Giuseppina Schiavone, Imen Chakroun, Ellie D’Hondt, Walter De Raedt, Jan Cornelis, Olivier Janssens, Sofie Van Hoecke, Stephan Claes, et al. 2018. Large-scale wearable data reveal digital phenotypes for daily-life stress detection.NPJ digi...

    2018

  52. [60]

    Paul Schmiedmayer, Anders Johnson, Narayan Schuetz, Lukas Kollmer, Paul Goldschmidt, Juan Delgado-SanMartin, Kelly W Zhang, Sriya D Mantena, Alexan- der Tolas, Samuel Montalvo, et al. 2026. Design and Rationale of the My Heart Counts Cardiovascular Health Study: a Large-Scale,...

    2026

  53. [61]

    James Truslow, Angela Spillane, Huiming Lin, Katherine Cyr, Adeeti Ullal, Edith Arnold, Ron Huang, Laura Rhodes, Jennifer Block, Jamie Stark, et al. 2024. Under- standing activity and physiology at scale: the Apple Heart & Movement Study. npj Digital Medicine7, 1 (2024), 242

    2024

  54. [62]

    Qingxin Wei, Jiaming Huang, Yi Gao, and Wei Dong. 2025. One Model to Fit Them All: Universal IMU-based Human Activity Recognition with LLM-assisted Cross-dataset Representation.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies9, 3 (2025), 1–22

    2025

  55. [63]

    Matthew Willetts, Sven Hollowell, Louis Aslett, Chris Holmes, and Aiden Doherty

  56. [64]

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. InInternational conference on machine learning. PMLR, 3319– 3328

    2017

  57. [65]

    Hao Xue and Flora D Salim. 2023. Promptcast: A new prompt-based learning paradigm for time series forecasting.IEEE Transactions on Knowledge and Data Engineering36, 11 (2023), 6851–6864

    2023

  58. [66]

    Hang Yuan, Shing Chan, Andrew P Creagh, Catherine Tong, Aidan Acquah, David A Clifton, and Aiden Doherty. 2024. Self-supervised learning for human activity recognition using 700,000 person-days of wearable data.NPJ digital medicine7, 1 (2024), 91

    2024

  59. [67]

    Cheng Zhang, Jinxin Lv, Jingxu Cao, Jiachuan Sheng, Dawei Song, and Tiancheng Zhang. 2025. Unravelling the semantic mysteries of transformers layer by layer. Comput. J.68, 9 (2025), 1237–1251

    2025

  60. [68]

    Hao Zhang, Zhan Zhuang, Xuehao Wang, Xiaodong Yang, and Yu Zhang. 2026. MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recog- nition.Advances in Neural Information Processing Systems38 (2026), 39334–39360

    2026

  61. [69]

    Maxwell A Xu, Girish Narayanswamy, Kumar Ayush, Dimitris Spathis, Shun Liao, Shyam A Tailor, Ahmed Metwally, A Ali Heydari, Yuwei Zhang, Jake Garrison, et al. 2025. Lsm-2: Learning from incomplete wearable sensor data.arXiv preprint arXiv:2506.05321(2025)

    2025 arXiv

  62. [70]

    Gupta, and Jingbo Shang

    Xiyuan Zhang, Ranak Roy Chowdhury, Rajesh K. Gupta, and Jingbo Shang. 2024. Large language models for time series: a survey. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence(Jeju, Korea)(IJCAI ’24). Article 921, 9 pages. doi:10.24963/...

    2024 doi

  63. [71]

    Xiyuan Zhang, Diyan Teng, Ranak R Chowdhury, Shuheng Li, Dezhi Hong, Rajesh K Gupta, and Jingbo Shang. 2024. Unimts: Unified pre-training for motion time series.Advances in Neural Information Processing Systems37 (2024), 107469– 107493

    2024

  64. [72]

    Yu Zhao, Rennong Yang, Guillaume Chevalier, Ximeng Xu, and Zhenxing Zhang

  65. [73]

    Built with Llama

    Jiahui Zhou, Dan Li, Lin Li, Zhuomin Chen, Shunyu Wu, Haozheng Ye, Jian Lou, and Costas J Spanos. 2025. Enhancing llm reasoning for time series classification by tailored thinking and fused decision.arXiv preprint arXiv:2506.00807(2025). 10 A Datasets We used seven datasets in...

    2025 arXiv

  66. [74]

    Mi Zhang and Alexander A Sawchuk. 2012. USC-HAD: A daily activity dataset for ubiquitous activity recognition using wearable sensors. InProceedings of the 2012 ACM Conference on Ubiquitous Computing. ACM, 1036–1043

    2012

  67. [78]

    Deep residual bidir-LSTM for human activity recognition using wearable sensors.Mathematical problems in engineering2018, 1 (2018), 7316954

    2018

  68. [80]

    sensor names MUST use one of these exact values: body_acc_x, body_acc_y, body_acc_z, gyro_x, gyro_y, gyro_z, total_acc_x, total_acc_y, total_acc_z↩→

  69. [81]

    beginning

    temporal MUST be EXACTLY one of: early | early_to_mid | mid | mid_to_late | late | full_window Replace any free-form phrase (e.g. "beginning", "around the middle", "toward the end") with the closest canonical label

  70. [82]

    DO NOT INVENT SENSOR BEHAVIOURS WHICH ARE NOT PRESENT

    pattern (OBSERVATION) MUST faithfully describe only what the evidence shows. DO NOT INVENT SENSOR BEHAVIOURS WHICH ARE NOT PRESENT. 21

  71. [83]

    inference MUST follow directly from the observations cited in its based_on field -- do not introduce biomechanical claims that are not grounded in those specific observations.↩→

  72. [84]

    [SYNTHESIS] MUST have a based_on field citing inference IDs (I1, I2, etc.)

  73. [85]

    Do NOT use any numbers, percentages, or timestep values in any text

  74. [86]

    Do NOT add any markdown formatting, code fences, headers, or bullet points outside the specified format

  75. [87]

    O<N_REGIONS>

    Observation IDs must be sequential: O1, O2, O3, ... O<N_REGIONS>

  76. [88]

    Inference IDs must be sequential starting from I1

  77. [89]

    Output ONLY the structured trace -- no preamble, no explanation, no commentary

  78. [90]

    early portion

    All output shall be in English only -- no Chinese or any other language ### PARAPHRASING RULES: Instead of numbers, use natural language: - Timesteps -> "early portion", "midway through", "toward the end" - Sensor values -> "sharp peak", "rapid oscillation", "sustained elevati...

  79. [2018]

    Statistical machine learning of sleep and physical activity phenotypes from sensor data in 96,220 UK Biobank participants.Scientific Reports8, 1 (2018), 7961

    2018

  80. [2022]

    Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems35 (2022), 23716–23736

    2022

  81. [2025]

    https://arxiv.org/abs/2510.15821

    Chronos-2: From Univariate to Universal Forecasting.arXiv preprint arXiv:2510.15821(2025). https://arxiv.org/abs/2510.15821

    2025 arXiv

  82. [2026]

    MantisV2: Closing the Zero-Shot Gap in Time Series Classification with Synthetic Data and Test-Time Strategies.arXiv preprint arXiv:2602.17868(2026)

    2026

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.