Skip to content

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

🕒 Published (v1): 2025-08-06 02:57 UTC · Source: Arxiv · Venue: ACL 2026 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

ZARA (Zero-training Activity Reasoning Agents) is a training-free agentic framework for Human Activity Recognition (HAR) from wearable motion sensors. It bridges the sensor-to-language gap by distilling labeled reference data into a pairwise statistical knowledge base and orchestrating three specialized LLM agents over class-conditional retrieved evidence. On 8 benchmarks, ZARA substantially outperforms all 10 baselines in parameter-frozen settings.

Problem

Conventional HAR models are trained end-to-end on fixed activity sets and sensor configurations; adapting to new users, activities, or hardware requires costly retraining. Direct LLM prompting with raw numerical time-series causes hallucinations and weak grounding because LLMs cannot intuit physical dynamics from raw numerical streams. Prior retrieval-augmented approaches in sensor HAR lack statistically verifiable motion priors, failing on fine-grained or long-tail activities without task-specific fine-tuning.

Method

ZARA decouples two information sources and routes them through a three-agent pipeline:

1. Offline Statistical Profiling (Universal Knowledge \(K\)). For every ordered activity pair \((a_i, a_j)\), extract a feature pool \(F\) (time-domain, frequency-domain, cross-channel statistics) from labeled windows \(x_a \in \mathbb{R}^{T \times C}\) and compute permutation-based importance scores via AutoGluon with cross-validated fold-weighted averaging. Store as \(K[(a_i,a_j)] = [(f_1,s_1),\ldots,(f_P,s_P)]\). New activities require only appending their pairwise profile—no retraining.

2. Class-Wise Multi-Sensor Retrieval. Maintain placement-specific FAISS vector stores \(\{D_\text{loc}\}\) embedding windows with a frozen TS encoder (Mantis by default, L2-normalized, inner-product similarity). For a query with candidate set \(A = \{a_1,\ldots,a_M\}\), perform class-conditional retrieval per placement, yielding ranked lists \(\{L^\text{loc}_m\}\). Fuse across placements via Reciprocal Rank Fusion: $\(\text{RRF}(d) = \sum_\text{loc} \frac{1}{k_\text{rrf} + r_\text{loc}(d)}, \quad k_\text{rrf}=60\)$ This ensures balanced recall for long-tail classes.

3. Hierarchical Multi-Agent Reasoning (four stages). - Feature Selector queries \(K\) over all candidates \(A\) to select \(n\) coarse discriminative features. - Evidence Pruning synthesizes retrieved class-wise evidence into a statistical comparison table (query value vs. class mean ± std), prunes distributionally mismatched activities, yielding refined set \(A'\). - Feature Selector re-queries \(K\) over \(A'\) for \(m\) fine-grained features. - Decision Insight compares query statistics against remaining candidates and outputs the final label with a natural-language rationale citing feature importance scores.

Key Contributions

  • Automated construction of a pairwise Activity Feature Importance Knowledge Base that translates implicit sensor statistics into verifiable linguistic priors without manual rule authoring.
  • First knowledge- and retrieval-driven multi-agent system for multi-sensor time-series classification that also generates evidence-backed natural-language rationales.
  • Class-conditional retrieval with RRF fusion ensures balanced recall across long-tail activities, closing the accuracy–F1 gap exhibited by all baselines.
  • Dynamic top-\(k\) class retrieval (cosine similarity against the database) decouples inference cost from activity library size for large-scale datasets.

Results

  • Cross-Subject (8 datasets): ZARA\(_\text{Gemini}\) achieves 81.6% avg accuracy / 81.4% macro F1, vs. best baseline UniMTS at 39.4% / 32.1% and second-best IMU2CLIP at 22.7% / 17.9%.
  • ZARA\(_\text{GPT}\) (GPT-4.1-mini): 77.5% / 77.2%; ZARA\(_\text{Qwen-80B}\): 73.4% / 73.0%; ZARA\(_\text{Qwen-30B}\): 71.0% / 70.2% — all exceed every baseline.
  • On "Easy" datasets (Opportunity, UCI-HAR, Shoaib), ZARA\(_\text{Gemini}\) reaches 92.5%, 90.0%, 97.1% accuracy respectively.
  • On "Hard" datasets (WISDM, DSADS), ZARA\(_\text{Gemini}\) scores 65.6% / 64.1% and 84.2% / 84.4%—substantially above all baselines while maintaining accuracy ≈ F1 alignment.
  • Cross-Dataset: Cross-dataset knowledge transfer is demonstrated across 5 dataset pairs (e.g., UCI↔USC, PAMAP2↔WISDM, MHealth↔WISDM), with cross-dataset knowledge achieving comparable or superior accuracy to in-dataset knowledge on several pairs.
  • Direct LLM prompting (HARGPT, Gemini Text/Table/Plot) averages 10–20% accuracy, confirming raw signal grounding is necessary.

Limitations

  • Requires a labeled support set per deployment domain to build \(K\) and populate vector stores; not purely zero-shot from a cold start.
  • Inference token cost scales with the number of activity candidates and agent stages (token usage statistics deferred to appendix).
  • Performance is coupled to backbone LLM numerical reasoning quality; weaker models show degraded F1 on fine-grained activities.
  • Cross-dataset evaluation is restricted to the intersection of shared sensor placements and activity labels, limiting coverage when hardware diverges significantly.
  • For hard benchmarks (e.g., USC-HAD with 12 classes), accuracy drops to ~60%, indicating residual difficulty with fine-grained activity separation.

Relevance to Harnesses / Meta-Harnesses

ZARA is a concrete example of a domain-specific multi-agent harness: it orchestrates three specialized LLM agents (Feature Selector, Evidence Pruning, Decision Insight) in a deterministic four-stage sequential pipeline, with each stage receiving structured context produced by the prior stage and the retrieval backbone. The offline knowledge base construction is a pre-harness compilation step analogous to index-building in RAG meta-harnesses, making the runtime pipeline inference-only. The design pattern—decompose a structured prediction task into specialist agents, each with a defined role, grounded inputs, and constrained output format—directly generalizes to meta-harness design for other modalities where raw data is not LLM-native. The class-conditional retrieval + RRF fusion sub-component is a reusable retrieval harness pattern for multi-index, multi-hypothesis retrieval problems.