Efficient and Adaptive Human Activity Recognition via LLM Backbones
Efficient and Adaptive Human Activity Recognition via LLM Backbones
Authors: Aleksandr Bredikhin, Philippe Lalanda, German Vega Date: 2026-05-12 Paper ID: openalex:2605.12019
Summary
This paper proposes repurposing pretrained LLMs as robust, adaptive temporal backbones for Human Activity Recognition (HAR) by replacing task-specific architectures. The authors introduce a structured convolutional projection to map inertial sensor data into the LLM latent space, while utilizing Low-Rank Adaptation (LoRA) to keep the backbone frozen. This methodology reduces training costs and data requirements while providing superior performance in cross-dataset and few-shot HAR scenarios. The approach effectively combines the local feature extraction of convolutional frontends with the long-range dependency modeling of LLM foundations.
Key Contributions
- Introduced a paradigm for leveraging frozen LLMs as generic temporal backbones for sensor-based Human Activity Recognition (HAR).
- Developed a structured convolutional projection that effectively bridges the modality gap between inertial time-series data and language model latent spaces.
- Demonstrated that LoRA-based fine-tuning of frozen LLM backbones achieves competitive accuracy with significantly reduced parameter count and improved data efficiency in few-shot scenarios.
Key Concepts
- structured-convolutional-projection: A technique for projecting multivariate inertial sensor data into the latent space of LLMs for downstream activity recognition.
Archivist Review
The paper proposes a specific modality-bridging technique (structured convolutional projection) that is sufficiently novel and reusable for time-series-to-LLM integration tasks to warrant a vault note. I have rejected no other candidates as none were provided. The review focused on identifying a reusable mechanism for cross-modal embedding that avoids generic architectural labels.
Approved Concepts
- Structured Convolutional Projection: It provides a novel methodology for bridging the modality gap between heterogeneous inertial time series data and frozen LLM backbones.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.12019
- paper_id
- 2605.12019
- paper_source
- openalex
- domain
- nlp
- tags
- nlp
- architectures
-
- datasets
-
- concept_slugs
- structured-convolutional-projection
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-15T06:16:39Z