Teaching smart homes to untangle who triggered each sensor

A smart home can detect that someone opened the refrigerator without knowing which resident wanted a snack. That missing identity matters if the eventual goal is to monitor an individual’s routines or personalize automation. Anonymous sensor readings become harder to interpret when several people move around together.

Xi Chen, Julien Cumin, Fano Ramparany and Dominique Vaufreydaz investigate this problem, known as data association. Their manuscript is an arXiv-posted author version, with a journal reference to Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 2026, 10(3), pages 1–37. November 4, 2026 is its original arXiv submission date, not an established journal publication date. Their system assigns events to resident tracks, including multiple residents for shared events. It separates event histories rather than establishing people’s real-world identities.

Learning linked decisions

The authors argue that attribution requires remembering everyone’s recent history, including residents who are not currently triggering sensors. A new kitchen event should be interpreted differently depending on whether another resident was recently occupying a chair elsewhere.

They formalize this process as a Markov decision process. The state combines the current event with previous assignments organized by resident. An action assigns that event, and the decision updates the histories used at the next step. Although the formal framework accommodates resident discovery, the implemented classifier uses a predefined maximum resident capacity.

Rather than learning through online trial and error, the models use behavior cloning: labeled sensor logs supply examples of correct decisions. One implementation trains a lightweight Transformer from scratch to predict a probability for each resident. Another fine-tunes Qwen3-4B-Instruct-2507 using low-rank adaptation. This language-model version receives textual descriptions of the house, sensors and resident histories, then generates a structured assignment.

The small Transformer can also use beam search, retaining several candidate assignment histories until later events help distinguish them. This is not used for the language-model policy, which generates assignments greedily. The main benchmark comparison uses greedy decoding for both policies; beam-search benefits are examined separately.

What the benchmarks show

The evaluation uses session-level training, validation and test splits in a 7:1:2 ratio across three datasets. CASAS contains two-person sessions; MARBLE includes one to four residents performing scripted activities; MuRAL includes two to four participants, naturalistic behavior and changing occupancy. Only ambient sensors are used.

An important qualification concerns the reported accuracy. Predictions cover whole sessions, but the default score evaluates overlapping windows of 16 events, relabeling identities locally by their order of appearance. It therefore measures local identity consistency rather than uninterrupted, session-long preservation of a particular identity. The researchers also measure correct tracking runs, sequence similarity and resident-count errors.

On MuRAL, the authors report accuracy of 77.3% for the fine-tuned language-model policy, compared with 72.8% for the small Transformer and 61.6% for LAHAR, a GPT-4o-powered baseline. On MARBLE, the corresponding figures are 82.4%, 79.2% and 73.0%. The standard language-model comparison excludes CASAS because the authors consider its densely deployed motion sensors difficult to describe faithfully in natural language.

Beam search produces additional, generally modest gains for the Transformer. On MuRAL, retaining eight candidate trajectories raises accuracy from 72.8% to 74.7% and lowers resident-count error from 0.437 to 0.312.

Where attribution still breaks

Shared events remain difficult. On MuRAL, the language-model policy recovers the correct number of residents for 41.3% of multi-resident events; most are under-assigned. That cardinality result should not be confused with proof that every participating identity was recovered correctly.

Mistakes also accumulate. In one illustrated MuRAL sequence, a delayed couch-sensor deactivation precedes an identity swap. The authors hypothesize that the model interpreted the timestamps too literally, rejecting a resident’s apparent movement to the bathroom. Subsequent decisions then inherited the mistaken history. Across the evaluation, accuracy generally falls as longer windows demand greater identity continuity.

Transfer between environments is uneven. Training on MARBLE yields 69.5% accuracy on MuRAL without additional fine-tuning, after replacing the environmental prompt. In the reverse direction, the authors report behavior resembling assignment of most events to one resident, despite superficially respectable accuracy. Removing sensor meanings and house-specific metadata also weakens the language model’s advantages.

A downstream activity-duration experiment suggests potential value for health monitoring, but assumes perfect activity recognition by using ground-truth activity labels. It is not a clinical validation, and substantial duration errors remain.

The authors describe the work as a technical foundation, not a deployment-ready system. Detailed identity annotations are costly, benchmark homes have unusually tailored sensor layouts, and performance in ordinary households with sparse sensors remains unvalidated. The evidence supports improved benchmark attribution—not a solved problem of reliably understanding everyone in a shared home.