Skip to main content
Back

UNSW's ZARA Lets Off-the-Shelf LLMs Classify Motion Without Retraining

A new agentic framework turns motion sensor signals into text-based reasoning problems, hitting 81.6% accuracy across eight benchmarks without any training.

Pro

  • ZARA is a training-free HAR framework using LLM agents guided by statistical priors, paper here.
  • Hits 81.6% accuracy across 8 benchmarks, beating UniMTS by over 40 points in parameter-frozen settings.
  • Decouples universal knowledge (pairwise feature importances) from local evidence (FAISS vector store of embeddings).
  • Hierarchical multi-agent pipeline: feature selection, class-wise retrieval, evidence pruning, decision insight with rationale.
  • Works across Gemini, GPT-4.1-mini, and Qwen3, showing gains come from the framework not the backbone.
  • Code available on GitHub; adding new activities requires only registering their statistical profile.

Human Activity Recognition from wearable motion sensors has long been stuck in a supervised rut. Every new activity, new user, or new device placement typically demands more labeled data and another round of training. A paper from researchers at UNSW Sydney and HKUST proposes a way out: ZARA, a training-free framework that lets off-the-shelf LLMs classify motion time-series by reasoning over statistically grounded evidence rather than raw numerical streams.

The signal-to-language gap

Existing HAR systems rely on task-specific deep networks optimized for fixed sensor setups and classes, so adapting to new users or hardware requires costly parameter optimization. Time-series foundation models help, but most still need a task-specific classification head trained on top. Contrastive approaches like UniMTS skip the classifier but struggle with fine-grained activities.

Handing raw accelerometer streams directly to an LLM does not work well either. Early attempts converting multi-channel signals into token sequences or images suffer from excessive token usage, significant information loss during discretization, and mediocre accuracy, because LLMs struggle to intuit physical dynamics from floating-point streams. Feeding a model a wall of numbers and asking whether someone is jogging or climbing stairs is a recipe for hallucination.

Comparison of HAR method families

Turning signals into verifiable priors

ZARA's core insight is to decouple two things that standard retrieval-augmented systems usually smash together. Universal Knowledge is a static reference registry storing pairwise feature-importance profiles, acting as a lookup table identifying which physical properties best separate specific activities, for example telling the agent that vertical acceleration variance is the critical metric to distinguish running from walking. Local Evidence is a vector database of raw-signal embeddings that serves as external memory, providing distributional grounding.