Parameter-Efficient Adaptation of Vision-Language Models for Wearable Artificial Intelligence: A Systematic Review of Resource-Constrained Deployment

This review adopted a structured narrative review methodology to identify, evaluate, and synthesize research on PEFT for large and giant VLMs and their applications in wearable AI. The methodology focused on PEFT techniques, multimodal foundation models, computational and memory efficiency, edge/on-device deployment, and resource-constrained wearable environments. The review covered studies published between 1 January 2020 and 31 December 2025. Initial database searches were conducted between 6 and 10 January 2026, citation tracking between 12 and 16 January 2026, and an updated database search between 1 and 5 February 2026. The review process comprised literature identification, duplicate removal, screening, eligibility assessment, quality evaluation, data extraction, and thematic synthesis.

2.1 Literature search strategy

The literature search was conducted in IEEE Xplore, Scopus, Web of Science Core Collection, ACM Digital Library, and ScienceDirect. The search strategy combined three conceptual groups: PEFT/adaptation techniques, VLMs/multimodal foundation models, and wearable/edge/resource-constrained AI. Search terms included "parameter-efficient fine-tuning", PEFT, LoRA, QLoRA, "low-rank adaptation", "adapter tuning", "prompt tuning", "prefix tuning", "vision-language model", VLM, "large vision-language model", "large multimodal model", "multimodal foundation model", "wearable AI", "wearable computing", "edge AI", "edge intelligence", "on-device AI", and "resource-constrained AI".

Database-specific Boolean search strings were developed to accommodate differences in database search syntax. For example, the Scopus search employed TITLE-ABS-KEY fields, Web of Science used TS, while IEEE Xplore, ACM Digital Library, and ScienceDirect used their respective advanced-search fields. Searches were limited to publications from 2020–2025 and, where supported, English-language publications.

Table 1 shows that the initial database searches retrieved 1,306 records: 214 from IEEE Xplore, 386 from Scopus, 297 from Web of Science, 241 from ACM Digital Library, and 168 from ScienceDirect. Backward and forward citation searching conducted from 12 to 16 January 2026 identified 74 additional records, resulting in 1,380 initially identified records. A database update conducted from 1 to 5 February 2026 retrieved 93 records, of which 41 were new after comparison with the existing review database.

Table 1. Database-specific search strategy and retrieval results

2.2 Eligibility, screening, and study selection

Studies were included when they were peer-reviewed journal articles or recognized international conference papers published between 2020 and 2025; investigated PEFT techniques such as LoRA, QLoRA, adapter tuning, prompt tuning, or related approaches; addressed VLMs, multimodal foundation models, or related architectures; and provided findings relevant to wearable AI, edge AI, on-device intelligence, resource-constrained computing, or efficient multimodal deployment. Studies focusing primarily on PEFT or VLM methodology without a direct wearable application were also retained when their findings had clear implications for resource-constrained or wearable deployment.

Studies were excluded when they focused exclusively on conventional full-model fine-tuning, lacked a relevant PEFT/VLM or deployment component, were editorials, tutorials, opinion articles, news or magazine publications, lacked sufficient technical information, were duplicates, fell outside the 2020–2025 period, or did not meet the predefined publication and language requirements.

Following consolidation of the retrieved records, 318 duplicates were removed, leaving 1,062 records for title and abstract screening. Of these, 781 were excluded, leaving 281 publications for full-text assessment. A total of 196 full-text publications were excluded because they were outside the review scope (72), lacked a PEFT component (43), lacked a relevant VLM/multimodal component (29), were not relevant to wearable/edge/resource-constrained AI (24), provided insufficient technical or methodological information (13), were editorial/tutorial/non-research publications (8), or represented duplicate/overlapping studies (7). Consequently, 85 studies were included in the initial qualitative synthesis.

The February 2026 update identified 41 new records. Of these, 18 proceeded to full-text assessment, and 7 additional studies met the eligibility criteria. Thus, the final qualitative synthesis comprised 92 studies.

2.3 Quality assessment and data extraction

The methodological quality of the included studies was assessed qualitatively using predefined criteria covering methodological rigor, technical reporting completeness, experimental validation, technical contribution, and relevance to wearable or resource-constrained deployment. Particular attention was given to the clarity of dataset descriptions, model architectures, PEFT configurations, training procedures, computational requirements, memory consumption, evaluation metrics, baseline comparisons, and reported limitations.

A structured data-extraction framework was applied consistently across the selected studies. Extracted information included publication characteristics, VLM architecture, PEFT technique, trainable-parameter proportion, datasets, application domain, adaptation strategy, computational and memory requirements, hardware platform, deployment architecture, edge/cloud/on-device strategy, inference characteristics, evaluation metrics, benchmark results, advantages, limitations, scalability considerations, and future research directions.

2.4 Citation tracking, search update, and reproducibility

Backward and forward citation tracking was conducted between 12 and 16 January 2026 to identify relevant studies not captured through the electronic database searches. References of eligible studies and relevant review papers were examined, while forward citation tracking was used to identify subsequent research citing influential publications. The 74 citation-derived records were subjected to the same screening criteria as database-retrieved records.

Because PEFT and VLM research is evolving rapidly, a supplementary database search was conducted between 1 and 5 February 2026 using the same search strategy and eligibility criteria. The update retrieved 93 records, of which 41 were new after comparison with the existing review database. This update resulted in seven additional eligible studies.

To enhance reproducibility, an audit trail was maintained containing the database name, search dates, update dates, exact database-specific search strings, search fields, publication-period restrictions, language restrictions, retrieved record numbers, duplicate-removal counts, screening decisions, full-text exclusion reasons, citation-search results, and final inclusion numbers. The exact search strings and search log should be provided as supplementary material where permitted by the target journal.

2.5 Data synthesis and analytical framework

The evidence was synthesized qualitatively because of substantial heterogeneity among the included studies in terms of VLM architectures, PEFT techniques, datasets, application domains, hardware platforms, experimental protocols, and evaluation metrics. Consequently, statistical meta-analysis was not considered appropriate.

The selected literature was organized into five thematic areas: (1) PEFT methodologies for VLM adaptation, (2) computational, parameter, and memory efficiency, (3) wearable, edge, and on-device deployment architectures, (4) multimodal wearable AI applications, and (5) open challenges and future research directions. Comparative analysis focused on parameter reduction, computational cost, memory requirements, adaptation performance, inference latency, deployment feasibility, scalability, and robustness.

Particular consideration was given to the practical constraints of wearable AI, including limited processing resources, memory capacity, energy consumption, inference latency, connectivity requirements, privacy, and the feasibility of local or edge-based model adaptation. This thematic synthesis enabled the review to move beyond a descriptive classification of PEFT methods by identifying how parameter-efficient adaptation can support the practical deployment of increasingly large multimodal models in wearable and resource-constrained intelligent systems.

2.6 Foundations of Parameter-Efficient Fine-Tuning

PEFT methods are designed to minimize the number of trainable parameters required during adaptation while preserving the representational capabilities of large pretrained models. Instead of modifying the entire parameter space, PEFT selectively updates compact parameter subsets or inserts lightweight trainable structures into existing architectures.

The theoretical foundation of PEFT stems from the observation that downstream adaptation often occupies a low-dimensional subspace within the original parameter manifold [15]. Early parameter-efficient adaptation strategies include adapter modules [23], prefix tuning [16], and prompt tuning [17]. These approaches demonstrated that competitive downstream performance can often be achieved by updating only a small fraction of model parameters, thereby motivating subsequent developments such as LoRA and QLoRA. Recent empirical evidence indicates that PEFT behavior in multimodal large language models is method- and architecture-dependent. Zhou et al. [24] evaluated multiple PEFT strategies across seven multimodal datasets and showed that the relative effectiveness of PEFT methods varies according to model architecture, adaptation-module location, training-data scale, and evaluation task. Their findings further indicate that adapter-based approaches can provide strong performance across several evaluated settings, highlighting the importance of empirical benchmarking rather than assuming that a single PEFT strategy is universally optimal.

2.6.1 Low-Rank Adaptation

PEFT has emerged as an important approach for adapting large VLMs by updating only a small subset of model parameters while preserving most pretrained weights. Among the proposed PEFT strategies are adapter-based methods, prompt tuning, prefix tuning, and LoRA techniques, which collectively aim to reduce the number of trainable parameters and the computational and memory requirements associated with model adaptation. This review examines these approaches from the perspective of wearable AI, where resource efficiency is a primary design consideration. LoRA has been widely adopted for adapting large language models and VLMs to downstream tasks, including multimodal question answering, industrial inspection, embodied AI, and edge intelligence. Within the reviewed literature, LoRA presents characteristics that may be relevant to smart glasses, augmented-reality headsets, and portable healthcare monitoring scenarios; however, suitability for specific wearable platforms requires validation on the target hardware.

The mathematical formulation of LoRA can be expressed as:

$W=W_0+B A$          (1)

where, W₀ denotes the original pretrained weight matrix, while B and A are trainable low-rank matrices. The output can therefore be represented as:

$h=\left(W_0+B A\right) x$           (2)

where, x denotes the input vector, and h represents the transformed output.

Recent studies demonstrate that LoRA-based adaptation achieves highly competitive performance in multimodal tasks, including image-text retrieval and visual reasoning [25], while typically reducing the number of trainable parameters by approximately 95% to over 99.9%, depending on the target architecture and selected rank configuration [15-18]. Variants such as QLoRA, AdaLoRA, and MoRA further optimize parameter allocation and quantization efficiency.

2.6.2 Adapter-based fine-tuning

Adapter methods insert lightweight neural modules between transformer layers. During training, only these adapter modules are updated while the base model remains frozen. Adapters offer modularity and facilitate rapid switching between downstream tasks.

Adapter-based tuning has potential relevance to personalized healthcare monitoring scenarios because task-specific modules can be trained while the backbone remains frozen; however, direct validation on wearable healthcare hardware remains limited. Adapters may reduce communication overhead in distributed settings because only compact modules require transmission between devices and servers.

Recent VLM-specific research also indicates that parameter-efficient adaptation need not always rely on additional trainable modules. Li et al. [26] demonstrated that selectively updating bias and normalization parameters can improve CLIP adaptation without introducing additional trainable modules. This finding broadens the PEFT design space and suggests that adaptation efficiency can also be achieved through selective modification of existing model parameters.

2.6.3 Prompt and prefix tuning

Prompt tuning adapts VLM behavior through learnable soft prompt embeddings prepended to input sequences. Prefix tuning extends this idea by optimizing trainable vectors inserted into transformer attention mechanisms.

These methods may be relevant to resource-constrained wearable systems. Multimodal prompt-tuning approaches extend conventional prompt-based adaptation by incorporating modality-specific information during parameter-efficient adaptation. For example, M²PT introduces multimodal prompt tuning for zero-shot instruction learning, addressing the limitation of approaches that treat modalities independently [27]. Such methods may be relevant to wearable multimodal systems because they provide a mechanism for adapting multimodal representations without updating the full model parameter space.

Multimodal PEFT has also been explored through query-based adaptation. Liang et al. [28] proposed Querying as Prompt, in which modality-specific queries are introduced as soft prompts to a frozen language model, providing an alternative mechanism for integrating multimodal information while limiting parameter updates.

2.6.4 Quantization-aware Parameter-Efficient Fine-Tuning

Quantization-aware PEFT combines parameter-efficient tuning with low-bit precision representations. QLoRA introduced a memory-efficient fine-tuning framework based on 4-bit quantized weights and reported competitive task performance [18].

For wearable systems with constrained memory and computational resources, quantized PEFT may improve adaptation and deployment feasibility by reducing model memory requirements; however, latency, thermal behavior, and energy consumption depend on the underlying hardware and quantization. Figure 2 compares representative PEFT methods in terms of trainable-parameter proportion, memory requirements, computational characteristics, and reported relevance to wearable deployment.

Figure 2. Comparison of Low-Rank Adaptation (LoRA), adapter tuning, prompt tuning, and QLoRA based on trainable parameters, memory requirements, computational complexity, and wearable deployment suitability

Table 2. Comparative characteristics of major Parameter-Efficient Fine-Tuning (PEFT) methods

As shown in Table 2, several low-rank and prompt-based PEFT methods use less than one percent of model parameters for adaptation, although the ratio varies substantially across adaptation strategies and model architectures. Such reductions in trainable parameters can be advantageous for memory-constrained adaptation; however, their implications for energy consumption depend on the hardware, workload, memory access patterns, and deployment configuration.

The quantitative comparison demonstrates that PEFT methods differ substantially in the degree and mechanism of resource reduction. Reported studies indicate that LoRA can reduce the number of trainable parameters by approximately 95% to more than 99.9%, depending on model architecture and rank configuration [15-18]. Prompt and prefix tuning can reduce the trainable parameter ratio to below 0.1%, making them attractive when adaptation-state size is the primary constraint. Adapter-based approaches generally require a larger trainable parameter fraction of approximately 1–5% but provide modularity that can be advantageous when multiple wearable tasks or personalized models must be supported.

QLoRA provides an additional efficiency mechanism by combining LoRA with 4-bit quantization. Its advantage therefore extends beyond the number of trainable parameters to the memory required for representing the frozen backbone [18]. However, parameter reduction should not be interpreted as a direct equivalent of inference-speed or energy reduction. Actual latency and energy consumption depend on hardware architecture, quantization support, memory bandwidth, model size, sequence length, and workload characteristics. Consequently, the reviewed evidence supports PEFT as a mechanism for reducing adaptation and memory costs, but it does not yet establish a universal ranking of PEFT methods in terms of wearable inference latency or battery consumption.

This distinction reveals an important limitation in the current literature. Existing studies frequently report parameter efficiency and model performance but provide fewer standardized measurements of battery consumption, thermal behavior, end-to-end latency, and sustained inference efficiency on actual wearable hardware. The absence of standardized hardware-aware measurements limits direct comparison among PEFT approaches and reinforces the need for wearable-specific benchmarking protocols that evaluate both model-level and system-level performance.

2.7 Key terminology

To ensure consistency throughout this review, the principal technical terms are defined in Table 3.

The definitions presented in the Table 3 establish a consistent conceptual framework for the terminology adopted throughout this review. These terms are subsequently used to evaluate and compare PEFT techniques according to their computational characteristics, deployment evidence, and potential relevance to wearable multimodal information systems.

The increasing scale of contemporary VLMs has created significant challenges for model adaptation and deployment. While full fine-tuning remains the conventional approach for transferring pretrained knowledge to downstream tasks, it becomes increasingly impractical as model sizes grow into the billions of parameters. Full fine-tuning generally imposes greater computational and memory requirements and may consequently increase energy demand, although the actual energy impact depends on the hardware and workload. These challenges are particularly problematic in wearable AI environments where computational resources are inherently limited. Consequently, PEFT has emerged as a promising paradigm for adapting large multimodal foundation models while reducing the number of trainable parameters and, in many settings, the associated adaptation memory and computational requirements.

Although recent studies have demonstrated parameter-efficient adaptation of multimodal models and increasingly efficient multimodal inference on mobile or edge hardware, direct evidence combining PEFT-enabled VLM adaptation with long-term evaluation on physical wearable platforms remains limited. In particular, standardized measurements of energy consumption, thermal behavior, latency, memory use, and adaptation performance across representative wearable hardware are not yet consistently reported. Consequently, the reviewed evidence supports the technical relevance of PEFT to resource-constrained multimodal systems but does not establish universal real-world deployment suitability.

Figure 3 presents a hierarchical taxonomy of PEFT techniques for VLMs, including adapter-based methods, prompt-based methods, LoRA methods, quantization-aware methods, and hybrid PEFT architectures. These approaches can be further grouped into additive-based techniques, selective parameter-updating methods, and reparameterization approaches. PEFT methods can be broadly categorized into additive-based techniques (adapters, prompt tuning, prefix tuning), selective parameter updating methods (BitFit, layer freezing, partial fine-tuning), and reparameterization approaches (LoRA, QLoRA, DoRA). These techniques can support efficient adaptation of vision encoders, language encoders, and cross-modal fusion modules while significantly reducing trainable parameters, computational cost, and memory requirements for deployment in resource-constrained environments such as wearable AI systems and edge devices. The taxonomy further highlights emerging hybrid PEFT architectures for multimodal VLMs and their deployment in wearable AI applications, including activity recognition, health monitoring, assistive systems, and edge intelligence.

The deployment of large VLMs in wearable environments requires computational architectures that balance model performance, latency, privacy, and resource constraints. Several deployment frameworks have emerged to support wearable multimodal intelligence.

Figure 3. Taxonomy of Parameter-Efficient Fine-Tuning (PEFT) techniques for Vision-Language Models (VLMs)

On-device deployment represents the most privacy-preserving architecture because all adaptation and inference processes occur locally on wearable hardware. This approach minimizes communication latency and eliminates dependence on external infrastructure. However, the severe computational constraints of wearable devices often limit the complexity of models that can be deployed. PEFT methods can improve the feasibility of adapting large models in resource-constrained environments by reducing the number of trainable parameters and associated adaptation costs. However, actual on-device deployment depends on the computational capability, memory capacity, energy budget, thermal characteristics, and software support of the target wearable platform.

6.1 Multimodal alignment

Recent studies have demonstrated that PEFT techniques, particularly LoRA, QLoRA, and adapter-based methods, can effectively adapt large VLMs while updating only a small fraction of model parameters. These approaches preserve most pretrained knowledge and achieve competitive performance across a wide range of multimodal tasks. However, the reviewed literature also indicates that parameter-efficient adaptation may not consistently preserve cross-modal alignment when models are deployed in potentially specialized wearable scenarios, such as personalized healthcare monitoring and industrial safety. Misalignment between visual and textual representations can reduce contextual reasoning accuracy under dynamic operating conditions. Future research should therefore investigate adaptive multimodal alignment mechanisms that continuously preserve semantic consistency while accommodating evolving user contexts and resource constraints.

6.2 Continual learning and model adaptation

Existing studies have shown that PEFT substantially reduces the computational cost of adapting large foundation models and supports incremental model updates without retraining the entire network. Nevertheless, most current approaches assume relatively static downstream tasks and are not specifically designed for continuously evolving wearable environments. As wearable devices generate non-stationary data streams and user behaviors change over time, PEFT-enabled models remain susceptible to catastrophic forgetting and performance degradation during continual adaptation. Future research should focus on lifelong multimodal learning frameworks that integrate PEFT with continual learning strategies capable of preserving previously acquired knowledge while adapting efficiently to new tasks and environmental changes.

6.3 Privacy and secure federated adaptation

Federated PEFT has emerged as a promising approach for enabling collaborative model adaptation without requiring centralized collection of sensitive user data. Although this paradigm significantly reduces direct data sharing, recent studies continue to report challenges related to communication overhead, gradient leakage attacks, model inversion risks, and limited scalability across heterogeneous wearable devices. These issues are particularly important for healthcare and other privacy-sensitive applications. Future research should therefore develop communication-efficient, privacy-preserving PEFT frameworks that integrate secure aggregation, differential privacy, and robust defense mechanisms while maintaining acceptable adaptation performance in distributed wearable ecosystems.

6.4 Explainability and trustworthy multimodal intelligence

Several studies have explored attention visualization and feature attribution techniques to improve the interpretability of large multimodal models. However, the decision-making processes of PEFT-enabled VLMs remain difficult to explain, particularly when deployed in safety-critical wearable applications. Limited interpretability may reduce user trust and hinder regulatory acceptance in domains such as healthcare and industrial monitoring. Future investigations should therefore focus on explainable PEFT architectures capable of providing transparent multimodal reasoning, uncertainty estimation, and human-interpretable explanations without significantly increasing computational complexity.

6.5 Energy-aware and hardware-constrained optimization

Recent advances, including QLoRA and other quantization-aware PEFT techniques, have significantly reduced memory requirements and computational complexity, thereby improving the feasibility of deploying large VLMs on resource-constrained devices. Despite these improvements, relatively few studies explicitly optimize adaptation strategies for battery consumption, thermal efficiency, hardware heterogeneity, and long-term operational sustainability in wearable systems. Differences in processor architectures, memory capacity, and communication capabilities continue to affect deployment performance across wearable platforms. Future research should investigate hardware-aware and energy-aware PEFT algorithms that jointly optimize model accuracy, computational efficiency, power consumption, and deployment scalability across diverse wearable devices.

6.6 Standardized benchmarking for wearable AI

Current benchmarking studies primarily evaluate PEFT techniques using conventional metrics such as accuracy, precision, recall, memory usage, and computational efficiency. While these metrics provide valuable insights, they do not fully capture the operational characteristics of wearable AI, including battery endurance, thermal behavior, communication latency, user personalization, and long-term adaptation performance. The absence of standardized evaluation protocols makes objective comparison among competing PEFT approaches difficult. Future research should establish comprehensive benchmarking frameworks specifically designed for wearable AI, incorporating both predictive performance and system-level deployment metrics to facilitate fair and reproducible evaluation. The major research gaps and corresponding future opportunities in PEFT-enabled VLMs for wearable AI are summarized in Table 7.