ZARA

ZARA

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

ACL 2026 Main Conference

Zechen Li   Baiyu Chen   Hao Xue   Flora D. Salim

University of New South Wales, Sydney

🌟 Overview

ZARA (Z̲ero-training A̲ctivity R̲easoning A̲gents) is a training-free, evidence-grounded LLM agent framework for motion time-series reasoning. It combines statistically grounded domain knowledge, multi-sensor retrieval-augmented evidence, and hierarchical multi-agent reasoning to classify motion sensor data with transparent, human-readable rationales—without fine-tuning or task-specific classifiers.

🔑 Key Features

• Training-Free HAR Inference: ZARA performs human activity recognition without task-specific fine-tuning or classifier training, enabling plug-and-play deployment in parameter-frozen settings.
• Evidence-Grounded Reasoning: Instead of relying on black-box projections, ZARA grounds LLM decisions in retrieved signal evidence and statistically derived feature priors.
• Pairwise Knowledge Base: ZARA builds an activity-pair knowledge base that captures discriminative motion properties as interpretable linguistic priors.
• Multi-Sensor Retrieval: Placement-specific retrieval and rank fusion provide class-balanced, query-relevant evidence across heterogeneous wearable sensors.
• Hierarchical Agentic Workflow: Specialized agents perform feature selection, evidence pruning, and final decision reasoning to produce concise, evidence-backed rationales.
• Strong Generalization: ZARA achieves state-of-the-art performance across 8 HAR benchmarks, with robust transfer across unseen subjects and heterogeneous sensor domains.

🚀 Code Examples

This release includes two end-to-end notebook examples:

• uci_notebook/uci_gemini.ipynb: UCI HAR inference with a single retrieval index, illustrating the non-RRF retrieval setting.
• shoaib_notebook/shoaib_gemini.ipynb: Shoaib HAR inference with placement-specific retrieval and reciprocal rank fusion (RRF).

Before running the Gemini-based inference notebooks, set your API key as an environment variable:

export GOOGLE_API_KEY="your_api_key"

Each dataset directory also includes preprocessing and feature-importance notebooks. Run preprocessing first, then feature-importance generation, and finally the Gemini inference notebook for the corresponding dataset.

Notice on feature-importance knowledge. The activity-pair knowledge base is constructed from AutoGluon feature importance and is used as a data-driven prior for ZARA. It should not be interpreted as an exhaustive list of all discriminative motion cues. Feature importance can be affected by the trained ensemble, validation split, feature correlation, and permutation behavior. As a result, some human-interpretable features may be under-ranked or omitted, especially when they are redundant with other features, useful only through interactions, or associated with rare motion patterns.

Practical ways to improve or audit the knowledge base include: increasing top_n when constructing pairwise prompts; aggregating importance across multiple random seeds or folds; combining permutation importance with model-native importance and simple statistical effect sizes. These steps can make the knowledge base more robust without changing the overall ZARA inference pipeline.

🌍 Citation

If you find this repository useful for your research, please cite our paper:

@article{li2026zara,
  title={ZARA: Training-Free Motion Time-Series Reasoning via Evidence-Grounded LLM Agents},
  author={Zechen Li and Baiyu Chen and Hao Xue and Flora D. Salim},
  journal={arXiv preprint arXiv:2508.04038},
  year={2026},
  url={https://arxiv.org/abs/2508.04038}
}

📄 License

This work is under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

📩 Contact

If you have any questions or suggestions, feel free to contact Zechen at zechen.li(at)unsw(dot)edu(dot)au.