DailyLLM: Context-Aware Log Generation

• DailyLLM is a context-aware system that integrates data from wearable devices to automatically produce detailed activity logs.
• It employs a modular three-stage pipeline combining data processing, feature extraction, and structured LLM prompting for human-readable narratives.
• DailyLLM achieves notable gains in semantic precision and inference speed, outperforming larger models in practical deployments.

DailyLLM is a lightweight, context-aware system for automatic activity log generation that integrates multi-modal sensor data from everyday wearable devices with LLMs to produce semantically rich and efficient activity narratives. Introduced in 2025, DailyLLM is the first framework to comprehensively combine location, motion, environmental, and physiological data solely from commonly available smartphone and smartwatch sensors to create high-fidelity user activity logs and summaries. Its core innovations include a structured prompting mechanism for the LLM, a robust feature extraction pipeline, and a deployment strategy that supports both high-performance machines and resource-constrained devices such as the Raspberry Pi. Experimental results demonstrate significant improvements in both semantic accuracy and inference speed compared to state-of-the-art baselines (Tian et al., 18 Jul 2025).

1. System Design and Architecture

DailyLLM is structured as a modular three-stage pipeline that transforms heterogeneous, time-series sensor data into detailed human-readable logs and activity summaries. The architecture comprises the following components:

• Data Collection and Processing: The system collects data from common smartphone and smartwatch sensors, including accelerometers, gyroscopes, GPS, Wi-Fi, Bluetooth, microphones, barometers, light, and basic physiological sensors (e.g., heart rate, GSR). Raw sensor streams are temporally aligned and transformed into timestamped JSON narratives with descriptive semantic labels. Enhanced processing steps, such as reverse geocoding and environmental categorization, expand raw signals with context labels (e.g., inferring "urban street" from GPS and barometric cues or light/temperature thresholds).
• Feature Extraction and Context Awareness: A dedicated engine extracts statistical and semantic features from each modality (see Section 2), yielding compact representations that align with human interpretable descriptors (such as "walking in a bright environment near a library at comfortable temperature").
• Structured Prompting & LLM Inference: The processed feature vectors are embedded into structured prompts with explicit roles, explanations, and output specifications. A lightweight LLM (1.5B parameters), fine-tuned with parameter-efficient methods, generates detailed log entries and temporal summaries. The LLM’s output is constrained by prompts that enforce fields such as time, location, activity category, environment, and scenario.
• Output Generation: Results are formatted into activity logs and higher-level summaries, with built-in mechanisms for anomaly detection and alerting (e.g., for unusual patterns or poor sensor coverage).

A schematic workflow is as follows:

[Sensor Collection] ↓ [Temporal Alignment & Semantic Tagging] ↓ [Multi-Modal Feature Extraction] ↓ [Structured Prompt Generation] ↓ [Lightweight LLM Reasoning] ↓ [Log & Summary Output]

2. Multi-Modal Feature Extraction

DailyLLM processes data across four dimensions to derive semantic context from low-level signals:

• Location Features: Combines GPS, barometer, Wi-Fi, and Bluetooth data. Reverse geocoding maps coordinates to structured addresses; signal metadata refines context (e.g., room-level distinction via wireless beacons).
• Motion Features: Aggregates IMU data (accelerometer, gyroscope, magnetometer) and computes statistical (mean, std, skewness, kurtosis, quantiles, entropy), frequency-domain (log energy per band, spectral entropy), and autocorrelation features. Axis-specific and inter-axis correlation coefficients are computed to capture movement directionality.
• Environment Features: Integrates light, ambient audio, and temperature. Sensor readings are mapped into semantic brackets (e.g., five light levels, seven noise brackets, three temperature zones). For audio, high-dimensional MFCCs and their derivatives are extracted to characterize ambient soundscapes.
• Physiology Features: Monitors physiological indicators (such as heart rate variability, GSR), retaining short-term statistics or the latest observed value per window.

This feature extraction condenses voluminous sensor data into interpretable, model-ready input while preserving context necessary for accurate semantic logging (Tian et al., 18 Jul 2025).

3. Structured Prompting Mechanism

DailyLLM employs a modular prompt framework to guide the LLM toward high-fidelity log generation and contextual summarization:

Prompt Section	Purpose
Data Introduction	Establishes expert role, e.g., "You are an expert in signal analysis"
Feature Explanation	Details sensor-derived attributes and context categories
Task Description	Describes the specific narrative or summary task
Feature Vectors	Inserts precomputed numerical/categorical features
Output Format	Specifies structured fields for LLM output

Structured prompts are dynamically composed with windowed sensor data (e.g., 2-10 min snapshots) and rich explanation to leverage both the LLM’s pretrained general world knowledge and its fine-tuned domain-specific capabilities. Explicit formatting requirements ensure consistency and facilitate downstream analysis.

This approach overcomes observed declines in accuracy from naive or free-form prompting strategies, delivering improved semantic richness and precise contextual inference.

4. Performance Evaluation and Comparison

DailyLLM demonstrates marked improvements over previous state-of-the-art (SOTA) log generation systems, both in content quality and computational efficiency:

• Semantic Precision: Using a 1.5B-parameter model, DailyLLM achieves a BERTScore precision of 81.65% in log generation—a 17% relative improvement over the SOTA baseline (AutoLife, 70B parameters, yielding ~65% precision).
• Inference Speed: The framework delivers a nearly 10-fold speedup over AutoLife and similar systems. Average timings for inference tasks on a personal computer (NVIDIA RTX 4090) were:
  • Activity prediction: 0.08s
  • Scene understanding: 0.13s
  • Log generation: 2.22s
  • 2-hour summarization: 1.63s (compared to 22.09s for baseline log generation).
• Resource Efficiency: DailyLLM can operate efficiently on low-resource platforms (such as Raspberry Pi 5 with ARM Cortex-A76, 8GB RAM) using memory-optimized batching, dynamic caching, and 6-bit quantized models. Summarization for a 2-hour window completes in under 4 minutes on such devices.

The table below contrasts DailyLLM and the SOTA baseline in key metrics:

Metric	DailyLLM (1.5B)	SOTA Baseline (70B)
BERTScore Precision	81.65%	~65%
Log Gen. Inference(s)	2.22	22.09
Model Size (params)	1.5B	70B

5. Deployment and Modality Integration

DailyLLM’s design emphasizes practical deployment, robust multi-modal coverage, and privacy:

• Sensor Economy: Utilizes only sensors found on commodity smartphones and smartwatches—no reliance on cameras or expensive hardware—maximizing deployability while minimizing privacy risks.
• Scalable Architecture: Supports parallel, dynamic batching and low-rank adaptation (LoRA) for efficient fine-tuning, enabling practical on-device or server-based deployment.
• Compact Feature Representation: Pre-inference feature extraction ensures the LLM processes only essential, information-rich input, reducing memory usage, latency, and storage overhead.

6. Applications and Future Directions

DailyLLM supports a spectrum of applications in ubiquitous and wearable computing, digital health monitoring, and life logging, where context-aware, interpretable activity logs are desirable for downstream analytics or personalized services. Notable properties include:

• Privacy Preservation: By eschewing camera-based sensing, DailyLLM protects user privacy and facilitates deployment in sensitive settings.
• Dynamic Summarization and Alerting: The system enables windowed log summarization and can issue anomaly alerts when sensor signatures deviate from learned patterns.
• Research Utility: The authors intend to release the underlying dataset and benchmark, supporting reproducibility and future studies in sensor-based activity monitoring.

Future work outlined in the original paper includes extending context windows for summarization, integrating additional sensor modalities, and enhancing privacy via further on-device optimization. There are also plans to generalize the dataset to capture greater diversity and granularity in daily activity contexts (Tian et al., 18 Jul 2025).

7. Significance and Comparative Perspective

DailyLLM establishes that fine-tuned, small-footprint LLMs—when combined with careful feature engineering and structured prompting—can outperform much larger models on practical, multi-modal activity log generation tasks. Its strongly modular design, resilience on resource-constrained edge devices, and robust multi-sensor integration position it as a notable development in the landscape of interpretable, context-aware behavioral analytics using LLMs. The system delivers not only higher semantic fidelity and efficiency, but also improved accessibility and privacy over previous SOTA approaches reliant on high-capacity hardware or limited modalities.

This comprehensive solution bridges raw sensor data and natural language, enabling next-generation life logging and digital health analytics with practical, scalable AI systems.