PaperlayerView exact revision on arXiv

Source-linked AI summary

ZARA: Training-Free Motion Time-Series Reasoning via Evidence-Grounded LLM Agents

Zechen Li, Baiyu Chen, Hao Xue, Flora D. Salim

arXiv:2508.04038v2cs.CLcs.CV

TL;DR

HAR commonly relies on task-specific training, while applying LLMs directly to numerical sensor streams can leave reasoning weakly grounded. ZARA builds textual motion priors and retrieves evidence to guide hierarchical agent reasoning; experiments report state-of-the-art generalization across unseen subjects and heterogeneous sensor domains. Its inference still requires a labeled retrieval database.

  • Problem

    HAR systems are commonly tied to fixed activity sets and costly parameter retraining, while direct LLM use on numerical time-series can lack grounding.

    Source: §1 · pp. 1–2

  • Method

    ZARA combines statistically profiled textual motion priors, placement-specific evidence retrieval, and hierarchical LLM agents for training-free inference.

    Source: §1 · p. 2, §3 · pp. 3–4

  • Results

    ZARA achieves state-of-the-art performance across cross-subject and cross-dataset settings, generalizing across diverse users and heterogeneous sensor domains.

    Source: §6 · p. 9

  • Takeaways & Limitations

    Retrieved evidence grounds transparent predictions, allowing off-the-shelf LLMs to be used without task-specific adaptation.

    Source: §6 · p. 9

    ZARA still requires a labeled retrieval database to ground its reasoning in representative evidence.

    Source: §7 · p. 9

Abstract

from arXiv · show

Motion sensor time-series are central to Human Activity Recognition (HAR), yet conventional approaches are constrained to fixed activity sets and typically require costly parameter retraining to adapt to new behaviors. While Large Language Models (LLMs) offer promising open-set reasoning capabilities, applying them directly to numerical time-series often leads to hallucinations and weak grounding. To address this challenge, we propose ZARA (Zero-training Activity Reasoning Agents), a knowledge- and retrieval-augmented agentic framework for motion time-series reasoning in a training-free inference setting. Rather than relying on black-box projections, ZARA distills reference data into a statistically grounded textual knowledge base that transforms implicit signal patterns into verifiable natural-language priors. Guided by retrieved evidence, ZARA iteratively selects discriminative cues and performs grounded reasoning over candidate activities. Extensive experiments on eight benchmarks show that ZARA generalizes robustly to unseen subjects and across datasets, demonstrating strong transferability across heterogeneous sensor domains. These results mark a step toward trustworthy, plug-and-play motion understanding beyond dataset-specific artifacts. Our code is available at https://github.com/zechenli03/ZARA.

1 Introduction

HAR systems commonly depend on task-specific training and struggle with training-free adaptation, semantic grounding, and transparent reasoning. ZARA combines statistical textual priors, retrieved evidence, and hierarchical agents; the paper reports strong generalization in parameter-frozen settings across unseen subjects and heterogeneous domains.

1 Introduction

  • Conventional HAR systems are typically trained for fixed sensor setups and activity classes, making adaptation to new users or hardware costly and limiting training-free use.Time-series foundation models may still require task-specific classification heads, while classifier-free contrastive models can lack fine-grained semantic grounding.Source: §1 · p. 1
  • Directly applying LLMs to sensor HAR through token or image conversions can incur high token use, information loss, and mediocre accuracy because physical dynamics are difficult to infer from raw numerical streams.The paper identifies statistically grounded language priors and retrieved signal evidence as the missing link for grounded training-free inference.Source: §1 · p. 2
  • ZARA profiles discriminative features for activity pairs, retrieves class-wise evidence, and uses specialized agents to filter features and candidates into evidence-backed explanations.The paper also highlights automated signal-to-text knowledge grounding and interpretable rationales as contributions.Source: §1 · p. 2
  • ZARA achieves state-of-the-art performance in parameter-frozen settings, generalizing across unseen subjects and heterogeneous domains.The authors attribute this transferability to injected motion priors.Source: §1 · p. 2

2 Related Work

Prior HAR approaches use time-series foundation models, cross-modal alignment, and LLM-based reasoning, but often retain task-specific training or lack fine-grained, verifiable grounding. ZARA is presented against this background of efforts to enable classifier-free and training-free recognition.

2 Related Work

  • General time-series foundation models such as Chronos, Moment, and Mantis provide learned representations, but typically still require downstream task-specific classifiers.Chronos uses scaling and quantization, Moment masked-value prediction, and Mantis contrastive pretraining.Source: §2 · p. 2
  • Cross-modal methods align motion with text or image embeddings, while UniMTS may lack semantic granularity and COMODO still uses task-specific training.ImageBind and IMU2CLIP align IMU data with vision-language spaces.Source: §2 · pp. 2–3
  • LLM-based HAR methods prompt on raw signals, generate captions, or use spatial metadata, but the cited approaches can incur token costs or information loss, require fine-tuning, or lack statistically grounded motion knowledge.The passage contrasts these approaches with the need for verifiable grounding for robust training-free generalization.Source: §2 · p. 3

3 Methodology

ZARA separates general pairwise motion knowledge from placement-specific retrieved evidence, then combines class-wise multi-sensor retrieval with hierarchical agent reasoning. Statistical profiles, rank fusion, and iterative feature selection support candidate filtering and evidence-grounded predictions.

3 Methodology

  • ZARA separates static pairwise feature-importance knowledge from local raw-signal evidence, using retrieval for user- or placement-specific grounding instead of adapting model weights.The universal registry identifies discriminative physical properties, while local evidence supplies distributional context.Source: §3 · p. 3
  • Offline profiling ranks interpretable time-, frequency-, and cross-channel features for each ordered activity pair using permutation-based importance estimates averaged across cross-validation folds.The resulting profiles can be instantiated for candidate subsets, and a new activity can be added by registering its profile against existing classes.Source: §3 · pp. 3–4
  • Placement-specific stores retrieve evidence from windows embedded by a frozen time-series encoder, restricting matches to distributionally aligned evidence for each sensor location.The stored embeddings are L2-normalized and indexed using FAISS IndexFlatIP.Source: §3 · p. 4
  • Class-wise retrieval supplies evidence for each candidate, and reciprocal rank fusion combines placement-specific rankings over time-synchronized sensor indices.The fusion favors windows salient across modalities, while per-class retrieval supports balanced recall for long-tail activities.Source: §3 · p. 4
  • Three specialized agent roles execute four stages: select coarse features, prune distributionally mismatched activities using statistical comparisons, select fine-grained features, then produce a label and evidence-grounded rationale.The final decision uses updated statistics for the refined candidate set.Source: §3 · p. 4

4 Experiments

ZARA is evaluated for cross-subject and cross-dataset generalization across eight HAR datasets. It outperforms the cross-subject baselines and shows that cross-dataset transfer depends on the source knowledge and data conditions.

4 Experiments

  • The two evaluation tiers test individual variation within a domain and transfer across heterogeneous sensor hardware and environments.For the cross-subject protocol, ZARA builds its knowledge base and retrieval index from seen subjects, then tests on a class-balanced held-out-subject split.Source: §4 · p. 5, §4.2 · p. 5, Figure 4 · p. 5

4.1 Datasets

  • The eight benchmarks are grouped into easy, medium, and hard difficulty levels, with three, three, and two datasets, respectively.Easy: Opportunity, UCI-HAR, and Shoaib; medium: PAMAP2, USC-HAD, and MHealth; hard: WISDM and DSADS.Source: §4.1 · p. 5

4.2 Cross-Subject Generalization

  • 81.6% average accuracy and 81.4% macro F1: ZARAGemini outperforms the strongest cross-subject baseline, UniMTS, with all tested ZARA backbones outperforming every baseline.The evaluation uses parameter-frozen settings.Source: §4.2 · p. 6
  • Cross-subject baselines show majority-class bias under distribution shift, while direct prompting fails without explicit reference grounding.The paper also reports that IMUGPT does not transfer from virtual pretraining to real-world benchmarks, and that ImageBind and IMU2CLIP are restricted to single-sensor inputs.Source: §4.2 · p. 6

4.3 Cross-Dataset Generalization

  • Cross-dataset knowledge outperforms in-dataset knowledge in 3 of 8 transfers, while in-dataset knowledge performs best in the other 5.The three transfers are UCI→USC, WISDM→PAMAP2, and WISDM→MHealth; retrieval evidence comes from the target domain while priors come from a source dataset.Source: §4.3 · pp. 6–7

5 Ablation Studies

Ablations show that ZARA's performance depends on its retrieval evidence, evidence pruning, and statistically grounded prior knowledge. The pipeline remains robust across retrieval embedders, with per-query latency varying by representation.

Removing Retrieval Reduces Performance

  • 81.6% to 71.8%: average accuracy falls when top-k retrieval is replaced with global class-wise distributions; ground-truth retention drops from 91.4% to 86.7%.The decline is especially pronounced where instance statistics diverge from global averages.Source: Figure 6 · p. 8, §5 · p. 8

Impact of Retrieval Embedder

  • With Moment-Small, ZARA beats the supervised baseline on 7/8 datasets in F1; with Moment-Large, it wins on both metrics on 7/8, and with Mantis it improves F1 on 4/8.Across representations, ZARA ranks first on 4 datasets and second on 7 despite using frozen embeddings only as retrieval anchors rather than training task-specific parameters.Source: §5 · pp. 7–8
  • Moment-Small is the fastest retrieval embedder, while DTW is substantially slower; Mantis is slower than Moment but retrieves more informative samples.The reported relative rankings capture method-level efficiency, though absolute latency varies with query data, database size, and hardware.Source: §5 · p. 8

Skipping Evidence Pruning Hurts

  • 81.6% to 68.2%: average accuracy falls without Evidence Pruning, although pruning retains 91.4% of ground-truth labels while typically narrowing queries to 2–3 candidates.Without pruning, the LLM reasons over the full candidate pool rather than focusing on the remaining hard negatives.Source: Figure 7 · p. 8, §5 · p. 8

No Prior Knowledge Fails

  • 81.6% to 63.4%: average accuracy falls without the pairwise feature-importance registry, and average ground-truth retention drops from 91.4% to 87.0%.Without statistical grounding, the selector can choose suboptimal features and overly coarse selections that cannot recover the lost accuracy.Source: §5 · pp. 8–9, Figure 8 · p. 9

6 Conclusions

ZARA combines statistical profiling, retrieval augmentation, and hierarchical reasoning to ground training-free HAR in explicit linguistic priors. Across eight benchmarks, it achieves state-of-the-art performance in cross-subject and cross-dataset settings, with transparent predictions grounded in retrieved evidence.

6 Conclusions

  • ZARA combines statistical profiling, retrieval augmentation, and hierarchical reasoning to translate sensor dynamics into explicit linguistic priors for verifiable, training-free inference.Labeled data serve as a reference registry rather than solely as training targets, allowing off-the-shelf LLMs to be used without task-specific adaptation.Source: §6 · p. 9
  • Across eight HAR benchmarks, ZARA achieves state-of-the-art performance in both cross-subject and cross-dataset settings.The reported results indicate transferable motion priors across user populations and heterogeneous sensor domains.Source: §6 · p. 9
  • Retrieved evidence grounds ZARA’s transparent predictions, which the authors present as a scalable and trustworthy path toward adaptive HAR in the wild.The conclusion frames this as a path toward adaptive HAR, not as a demonstrated deployment outcome.Source: §6 · p. 9

7 Limitations

ZARA occupies a distinct design space, limiting direct architectural comparisons. Its inference still depends on labeled retrieval data, and its agentic workflow costs more than simpler unimodal pipelines.

7 Limitations

  • Direct architectural comparisons are limited because ZARA combines domain knowledge, retrieval augmentation, and agentic reasoning in a distinct design space.SensorLM, LLaSA, and RelCon are excluded because their input modalities do not match the synchronized, multi-location motion benchmarks.Source: §7 · p. 9
  • Although parameter-frozen at inference, ZARA requires a labeled retrieval database to ground reasoning in representative evidence.The support set serves as a reference registry and does not update model weights.Source: §7 · p. 9
  • ZARA’s agentic workflow has higher computational cost than simpler unimodal pipelines; future work will explore more efficient designs and domain-adapted LLMs.The stated aim is to reduce dependence on external references while preserving transparency and grounded reasoning.Source: §7 · p. 9

A Appendix

The appendix supplies implementation and evaluation details that complement the main-text aggregate results.

A Appendix

  • The appendix covers baseline models, dataset and preprocessing details, per-dataset token usage, per-class benchmark performance, and the full prompts used at each ZARA reasoning stage.The per-dataset and per-class results complement the main-text aggregate results.Source: A Appendix · p. 13

A.1 Baselines

The appendix documents how baseline methods were reproduced or adapted for training-free HAR evaluation, alongside cross-subject dataset statistics.

A.1 Baselines

  • Implementation details describe how the study’s baseline methods were reproduced or adapted for training-free HAR evaluation.The appendix also reports cross-subject dataset statistics.Source: §A.1 · p. 13, Table 3 · p. 14
  • HARGPT is tested with raw numerical input downsampled to 10 Hz and with plotted signals rendered as sensor subplots in a composite figure.The visual-input baseline uses GPT-4o-mini.Source: §A.1 · p. 13
  • Gemini-2.0-Flash is evaluated on raw sequences, plotted signals, and Markdown tables to assess ZARA’s modules beyond the base model.The baseline uses the same LLM backbone as ZARA.Source: §A.1 · p. 13
  • ImageBind and IMU2CLIP are evaluated one sensor placement at a time, with repeat padding and interpolation to meet their fixed input-window lengths.ImageBind uses 6 × 2000 windows and reports the best result across placements; IMU2CLIP uses 6 × 1000 windows and reports the best result for each dataset.Source: §A.1 · p. 13
  • NormWear is evaluated for activity recognition using its recommended prompt together with candidate activity labels.The model is designed to extract generalized representations from multivariate wearable signals.Source: §A.1 · p. 13
  • IMUGPT’s synthetic training data are generated from textual activity descriptions converted to 3D motion and virtual IMU streams; evaluation uses DeepConvL-STM without supervised distribution calibration.The calibration stage is excluded because it depends on labeled downstream data.Source: §A.1 · p. 13
  • UniMTS is evaluated with released checkpoints and text-label matching in its training-free inference setting, following the official implementation.Its unified pretraining aligns motion signals with LLM-enriched text descriptions for classifier-free recognition.Source: §A.1 · p. 13

A.2 Retrieval Strategies

ZARA combines ranked retrieval within sensor placements with cross-placement Reciprocal Rank Fusion. The appendix describes DTW, Moment, and Mantis as retrieval backbones with distinct preprocessing and representation choices.

A.2 Retrieval Strategies

  • ZARA ranks retrieval candidates within each sensor placement, then aggregates the ranked lists across placements using Reciprocal Rank Fusion.Source: §A.2 · p. 14
  • DTW retrieval z-score normalizes six sensor channels, negates distances as similarity scores, and returns the top-k candidates.Source: §A.2 · p. 14
  • Moment averages per-channel embeddings for retrieval, whereas Mantis concatenates channel-wise embeddings into a unified representation.The appendix evaluates Moment-Small and Moment-Large; Mantis rescales sequences to length 512 and produces 256-dimensional embeddings per channel.Source: §A.2 · p. 14

A.3.1 Cross-Subject Generalization

The supplementary sections detail cross-subject and cross-dataset evaluation protocols, token use, ablations, feature construction, and inference prompts. They report per-class performance patterns, component effects, and retrieval latency alongside those methodological details.

A.3.1 Cross-Subject Generalization

  • Cross-subject evaluation samples balanced, non-overlapping windows by activity and subject, using identical subsets across baselines.Source: §A.3.1 · p. 14

A.3.2 Cross-Dataset Generalization

  • Cross-dataset transfer moves feature guidance only, retrieves from the target dataset, and aligns shared activities and sensor placements for comparable evaluation.Source: §A.3.2 · p. 16

A.4 Token Usage Breakdown

  • Token use is dominated by early agent stages, input tokens exceed output tokens, and prompts are shorter than direct raw-signal linearization approaches.Source: Table 5 · p. 17, §A.4 · p. 17

A.5 Ablation

  • Evidence pruning substantially improves performance while preserving high-quality candidates, and prior knowledge is critical for narrowing candidates and distinguishing activities.Source: §A.5 · pp. 17–18
  • Across eight datasets, DTW is slowest, Moment-Small fastest, and Mantis offers a speed–retrieval-quality trade-off; latency varies with window length, channel count, and database size.Source: Table 9 · p. 18, §A.5 · p. 18

A.6 Predefined Features

  • The predefined feature pool combines interpretable time- and frequency-domain statistics across each sensor’s six axes and two magnitude channels.Source: §A.6 · p. 18, Table 10 · p. 19

A.7 Prompt Template

  • Inference prompts coordinate feature selection, evidence pruning, shortlist refinement, and final class decisions grounded in query features and class distributions.Source: §A.7 · pp. 18, 21, 22, 23

A.5 Ablation

  • ZARA consistently achieves strong per-class performance, while several baselines have substantially lower F1 than accuracy, indicating dominant-class bias.Source: §A.5 · p. 18, Figure 9 · p. 19

Related papers

Loading 2508.04038v2…