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

ZARA is a training-free, retrieval-augmented agent framework that converts motion sensor data into verifiable textual priors to enable robust, open-set human activity recognition without requiring parameter retraining.

The Gist

Imagine you are trying to teach a very smart, well-read librarian (a Large Language Model, or LLM) how to recognize different human movements just by looking at numbers from a smartwatch.

The problem? If you just hand the librarian a raw spreadsheet of numbers (like "acceleration: 0.5, -0.2, 0.8..."), they get confused. They might guess "dancing" when the person is actually "walking," or they might make up a completely fake activity because they are trying too hard to be creative. This is called "hallucination."

ZARA is a new system that fixes this by giving the librarian a detective's toolkit instead of just a raw data dump. It allows the AI to recognize human activities without needing to be retrained on new people or new devices.

Here is how ZARA works, broken down into simple analogies:

1. The Problem: The "Black Box" vs. The "Detective"

Most current methods are like a black box. You feed it data, and it spits out an answer. If you show it a new type of watch or a new person, the black box breaks because it was only memorized for the specific people and watches it saw during training.

ZARA is like a detective. It doesn't just guess; it investigates. It asks: "What specific clues in this data prove this person is running and not walking?"

2. The Three Pillars of ZARA

ZARA uses three main tricks to act like a super-detective:

A. The "Cheat Sheet" (Statistical Knowledge)

Imagine you want to explain the difference between walking and running to someone who has never seen them.

• Old Way: You show them a video of a person running and a person walking.
• ZARA's Way: You give them a Cheat Sheet that says: "Running has much higher 'vertical bounce' (up and down movement) than walking. Walking is smoother."

ZARA automatically creates these "Cheat Sheets" (a textual knowledge base) by analyzing thousands of past movements. It turns boring numbers into clear, human-readable rules (e.g., "If the arm swings fast and the heart rate is high, it's likely jogging"). This gives the AI a solid foundation of facts before it even looks at the new data.

B. The "Reference Library" (Retrieval)

When the AI sees a new movement, it doesn't just guess. It goes to its Reference Library.

• It asks: "I see a movement that looks like jogging. Do I have any past examples of jogging from this specific type of watch to compare it against?"
• It pulls up the most similar past examples (Evidence).
• It compares the new movement to these specific examples to see if they match.

This is like a chef tasting a new soup and comparing it to a specific recipe they have on hand, rather than just guessing the ingredients based on a vague memory.

C. The "Team of Specialists" (Agentic Reasoning)

ZARA doesn't rely on one big brain. It uses a team of specialized agents (like a courtroom jury) to make the final decision:

1. The Feature Selector: Looks at the "Cheat Sheet" and says, "Okay, for this specific comparison, the most important clue is the 'vertical bounce'."
2. The Evidence Pruner: Looks at the "Reference Library" and says, "We can rule out 'sleeping' and 'eating' immediately because the data doesn't match those patterns at all. Let's focus only on 'walking' and 'jogging'."
3. The Decision Maker: Takes the remaining clues, compares them to the library examples, and makes the final call: "It's jogging!"

Crucially, this team writes down their reasoning. Instead of just saying "Jogging," they say: "We chose Jogging because the vertical bounce was 80% higher than walking, and the arm swing matched our library examples for jogging." This makes the AI trustworthy.

3. Why is this a Big Deal?

• No Re-training: Usually, if you want an AI to recognize a new activity (like "yoga") or work on a new person, you have to spend weeks retraining the computer. ZARA just needs to add a new "Cheat Sheet" entry for yoga. It works instantly.
• Works Everywhere: Because it relies on general rules (physics of movement) rather than memorizing specific people, it works well even if the sensor is on a different part of the body or a different brand of watch.
• Trustworthy: In medical or safety situations, you can't just trust a "black box." ZARA explains why it made a decision, which is vital for doctors or safety systems.

Summary Analogy

Think of Old AI as a student who memorized a specific textbook. If the exam questions change slightly, they fail.

Think of ZARA as a seasoned detective.

1. They have a file of rules (Knowledge) about how crimes (movements) usually happen.
2. They have a database of past cases (Retrieval) to compare against.
3. They interview witnesses (Agents) to narrow down suspects.
4. They write a report explaining exactly why they caught the criminal.

ZARA allows computers to understand human movement as naturally as a human detective, without needing to go back to school every time a new person walks into the room.

Technical Summary

1. Problem Statement

Human Activity Recognition (HAR) from wearable motion sensors is critical for digital health and adaptive interfaces. However, current state-of-the-art approaches face three major barriers to scalable deployment:

• Poor Generalization: Existing deep learning models are heavily supervised and require costly retraining (parameter optimization) to adapt to new users (cross-subject) or different hardware setups (cross-domain).
• Limited Training-Free Adaptation: While time-series foundation models (e.g., Moment, Mantis) offer transferable representations, they still require task-specific classification heads. Contrastive models often struggle with fine-grained activity distinction in parameter-frozen settings due to weak semantic grounding.
• Lack of Interpretability: Most methods produce categorical predictions without transparent reasoning, limiting trust in safety-critical scenarios.

Directly applying Large Language Models (LLMs) to raw numerical time-series has also failed due to "hallucinations," excessive token usage, and the inability of LLMs to intuit physical dynamics from raw streams without explicit grounding.

2. Methodology: The ZARA Framework

The authors propose ZARA (Zero-training Activity Reasoning Agents), a novel agentic framework that enables training-free inference by bridging the gap between implicit sensor signals and explicit natural language reasoning. ZARA utilizes a Knowledge- and Retrieval-Augmented Generation (RAG) approach, decoupling universal knowledge from local evidence.

The framework consists of three synergistic components:

A. Offline Statistical Profiling (Global Priors)

Instead of embedding sensor priors into model weights, ZARA constructs a Pairwise Activity Feature Importance Knowledge Base ($K$).

• Process: For every pair of activities (e.g., Walking vs. Running), the system extracts low-cost, interpretable statistical features (time-domain, frequency-domain, cross-channel) from labeled data.
• Mechanism: Using permutation-based feature ranking (via AutoGluon), it calculates an importance score for each feature in distinguishing specific activity pairs.
• Output: A structured textual registry translating implicit signal statistics into verifiable linguistic priors (e.g., "Vertical acceleration variance is the critical metric to distinguish Running from Walking"). This allows the system to accommodate new activities simply by registering their profiles without retraining.

B. Class-Wise Multi-Sensor Retrieval (Local Evidence)

To ensure the LLM reasons over relevant local data, ZARA employs a retrieval backbone.

• Placement-Specific Stores: Vector databases are maintained for specific sensor locations (e.g., wrist, ankle) using embeddings from a frozen time-series foundation model (e.g., Mantis).
• Class-Conditional Retrieval: For a query, the system retrieves top-$k$ evidence for each candidate activity class independently from the relevant sensor databases.
• Rank Fusion: Rankings from different sensor modalities are fused using Reciprocal Rank Fusion (RRF) to create a unified, time-synchronized evidence set. This ensures balanced recall, even for long-tail activities.

C. Hierarchical Multi-Agent Reasoning

ZARA orchestrates a four-stage workflow using specialized LLM agents:

1. Feature Selector Agent: Queries the Knowledge Base ($K$) to identify coarse-grained discriminative features for the current candidate set.
2. Evidence Pruning Agent: Uses the retrieved evidence and statistical distributions (mean $\pm$ std) to filter out implausible activities, narrowing the hypothesis space.
3. Refined Feature Selector: Re-engages on the pruned set to select fine-grained features for resolving subtle ambiguities.
4. Decision Insight Agent: Analyzes the final statistics and retrieved evidence to produce the final prediction and a human-readable, evidence-backed rationale.

3. Key Contributions

1. Signal-to-Text Knowledge Grounding: An automated method to distill motion time-series into a pairwise textual knowledge base, enabling LLMs to perform verifiable reasoning in a parameter-frozen setting.
2. Agentic Framework for Interpretable HAR: The first knowledge- and retrieval-driven system for multi-sensor time-series classification that generates concise, evidence-backed rationales, enhancing trust in automated decision-making.
3. Strong Training-Free Generalization: Demonstrated state-of-the-art performance in parameter-frozen settings, showing robust generalization across unseen subjects and heterogeneous sensor domains without task-specific fine-tuning.

4. Experimental Results

The authors evaluated ZARA on 8 diverse HAR datasets (ranging from Easy to Hard difficulty, e.g., UCI-HAR, PAMAP2, WISDM, DSADS) against 10 established baselines, including text-based LLMs, multimodal LLMs, and pretrained foundation models (UniMTS, ImageBind, Mantis).

• Cross-Subject Generalization: ZARA consistently outperformed all baselines. The best variant (ZARA-Gemini) achieved an average accuracy of 81.6% and macro F1 of 81.4%, significantly surpassing the strongest baseline (UniMTS at ~39% Acc).
• Cross-Dataset Generalization: ZARA demonstrated strong transferability when using knowledge derived from a source dataset to infer on a target dataset with different hardware. Notably, knowledge from diverse sources (e.g., WISDM) generalized better to smaller datasets than local knowledge, suggesting the capture of transferable motion priors.
• Ablation Studies:
  • Retrieval: Removing the retrieval module dropped accuracy from 81.6% to 71.8%, proving the necessity of local evidence anchoring.
  • Pruning: Removing the Evidence Pruning agent caused a sharp drop to 68.2%, highlighting the value of narrowing the candidate space.
  • Prior Knowledge: Disabling the knowledge base caused a massive drop to 63.4%, confirming that general LLMs cannot reliably infer discriminative motion properties without statistical grounding.

5. Significance and Impact

ZARA represents a paradigm shift in HAR from training-intensive pipelines to plug-and-play, evidence-grounded reasoning.

• Trustworthiness: By providing transparent, natural-language rationales grounded in retrieved statistical evidence, ZARA addresses the "black box" issue of deep learning models, making it suitable for safety-critical applications.
• Scalability: The training-free nature eliminates the need for costly retraining when deploying to new users or devices, solving a major bottleneck in real-world HAR deployment.
• Generalization: The ability to transfer motion priors across heterogeneous sensor domains suggests a path toward universal activity understanding that is not bound by specific dataset artifacts.

In conclusion, ZARA successfully translates implicit sensor dynamics into explicit linguistic priors, enabling off-the-shelf LLMs to perform robust, interpretable, and generalizable human activity recognition without parameter updates.