MonitorVLM: Multidomain Monitoring Overview

• MonitorVLM is a research pattern that defines diverse monitoring pipelines for both ARINC 664 virtual link communications in avionics and vision-language model-based anomaly detection in various domains.
• It integrates structured context, selective multimodal reasoning, and explicit memory mechanisms to assess consistency, drift, and safety in real time or near-real time.
• Practical implementations range from SystemC-based simulation in integrated modular avionics to VLM-driven monitoring in scientific labs, driver cabins, mining safety, continual learning, and robotics.

MonitorVLM is an overloaded designation in the arXiv literature rather than a single canonical system. In one line of work, it denotes Virtual Link Monitoring for ARINC 664/AFDX communication inside integrated modular avionics (IMA). In later work, it denotes or motivates vision-language-model-based monitoring for scientific laboratories, driver monitoring, mining safety, embodied robotics, continual learning, and even the monitoring of VLM perceptual competence itself (Aloui et al., 2017, Lin et al., 4 Jun 2025, Wu et al., 4 Oct 2025, Yuan et al., 8 Jun 2026, Hyeon-Woo et al., 2024).

1. Terminology and research scope

A central source of ambiguity is the acronym VLM. In the avionics paper, VLM means Virtual Link Monitoring and refers to ARINC 664 communication channels (Aloui et al., 2017). In most later papers, VLM means Vision-LLM, and “MonitorVLM” refers to a monitoring pipeline that uses multimodal reasoning over images, video, prompts, or structured context (Lin et al., 4 Jun 2025).

Usage of “MonitorVLM”	Monitored object	Representative source
Virtual Link Monitoring in IMA	ARINC 664/AFDX traffic and partition timing	(Aloui et al., 2017)
Visual anomaly detection in robot laboratories	First-person images of scientific workflow steps	(Lin et al., 4 Jun 2025)
Driver monitoring concept	In-cabin driver state from video frames and prompts	(Cañas et al., 15 Mar 2025)
Mining safety violation detection	Surveillance video frames mapped to clause-specific violations	(Wu et al., 4 Oct 2025)
Online video anomaly detection with LLM/VLMs	Streaming surveillance anomalies with memory and scoring queue	(Yang et al., 24 Oct 2025)
Monitoring online TFCL	Distribution drift and shift etiology in non-stationary streams	(Yuan et al., 8 Jun 2026)
Monitoring VLM perceptual competence	Color, shape, and semantic sensitivity of deployed VLMs	(Hyeon-Woo et al., 2024)
Crop monitoring and task planning	Long-horizon plant inspection with tool-calling VLMs	(Cuaran et al., 17 Jan 2026)
Embodied visual tracking recovery	Target-loss events and recovery plans in dynamic scenes	(Wu et al., 27 May 2025)

The shared conceptual thread is not a fixed architecture, but the use of a monitoring subsystem that sits beside an operational process and evaluates consistency, anomalies, drift, or safety conditions from privileged observations. The monitored process, however, differs sharply across domains: packetized avionics communication, robotic workflows, driver behavior, mining surveillance, continual learning streams, and VLM internals.

2. Avionics origin: MonitorVLM as Virtual Link Monitoring

In "Embedded real-time monitoring using SystemC in IMA," MonitorVLM is a Virtual Link Monitoring for ARINC 664/AFDX networks in IMA capability implemented as a SystemC simulator running inside the avionics switch modules (ASM) (Aloui et al., 2017). The ASM is described as a “privileged place” because “the CPU only manages message traffic and have available time,” which allows co-located monitoring without disturbing end systems. The simulator observes ARINC 664 traffic and compares it against a timed model derived from the ARINC 653 major frame (MAF), partition windows, and ARINC 664 Virtual Link configuration, notably BAG, frame size, and jitter. Each partition is assigned a dedicated Virtual Link, and the monitor maps VLs to partitions to verify temporal sequencing and payload coherence.

The implementation relies on SystemCASS, adopted to avoid nondeterministic dynamic event scheduling and to provide static, deterministic scheduling with CFSM models at CABA abstraction and a single clock domain (Aloui et al., 2017). The paper specifies three SystemCASS process types—Transition, Moore Generation, and Mealy Generation—and reports that dynamic memory allocation was removed by using a static gcc and replacing runtime dynamic allocations. SystemCASS performs static dependency analysis at startup and creates a fixed schedule for process execution throughout simulation time.

The monitored properties are narrowly defined. Temporal consistency checks whether communication occurs at the expected time according to the ARINC 653 schedule and ARINC 664 timing, including detection of missing communications and unexpected communications. It also verifies that the execution order of partitions is respected, with the “sequencing number of partitions” referring to partition execution over the MAF rather than ARINC 664 frame sequence numbers. Data consistency compares consecutive values $T$ and $T+1$ by a physical variation law for applications such as GPS, Speed, and Angle; incoherent evolution raises an alarm. By contrast, the monitor does not cover ARINC 664 Part 7 redundancy (A/B links), frame sequence number checks, CRC error handling, formal bandwidth policing, or analytical performance bounds. Validation is demonstrator-oriented rather than quantitative: the transmission side uses POK OS applications and a SystemC End System, and a demonstrator is planned with two QorIQ T2080 boards (PowerPC E6500) (Aloui et al., 2017).

3. Context-dependent visual monitoring in laboratories and driver cabins

In "A VLM-based Method for Visual Anomaly Detection in Robotic Scientific Laboratories," MonitorVLM is a VLM-based monitoring pipeline for detecting visual anomalies in robotic scientific experiment workflows (Lin et al., 4 Jun 2025). The method uses first-person RGB images resized to 640×480 together with structured textual context. Its formalization is

$$\text{judgeAnomaly}(x, I, \phi(S, C, C_d)) \rightarrow y,$$

where $x$ is the image, $I$ is fundamental experiment context, $S$ is the ordered set of steps, $C$ the detection target, $C_d$ the fine-grained anomaly description, $\phi$ the information control function, and $y \in \{0,1\}$. The paper introduces four progressively informative prompt configurations: Level 1 uses only Experiment Context; Level 2 adds Stage Description; Level 3 adds Detection Content; and Level 4 adds Anomaly Label Description. The benchmark comprises 1001 first-person images (640×480), with 501 normal and 500 abnormal samples, organized into 15 discrete stages and 20 monitoring points. On this benchmark, GPT-4o improves from ACC 41.6% at Level 1 to 79.2% at Level 4, with FPR 16.8% and MDR 3.0% at Level 4; Qwen2.5-VL-72B-Instruct improves from 41.6% to 61.4%, also reaching MDR 3.0% at Level 4, though with FPR 35.6% (Lin et al., 4 Jun 2025). The paper’s central result is that richer workflow context materially improves anomaly detection.

A related but more exploratory formulation appears in "Exploration of VLMs for Driver Monitoring Systems Applications," where MonitorVLM is a vision–language–model-driven Driver Monitoring System (DMS) concept (Cañas et al., 15 Mar 2025). Its scope includes drowsiness via observable proxies such as yawning, distraction such as texting/typing, phone calls, reaching/adjusting hair, and drinking, as well as coarse gaze direction, head pose, and general alertness. The experiments use RGB frames from the Driver Monitoring Dataset (DMD) and evaluate zero-shot and one-shot prompting with Idefics2. The paper reports qualitative findings rather than formal quantitative metrics: downward gaze was easier to identify, left/right gaze was inconsistent, one-shot prompting could correctly identify behaviors such as “adjusting his hair while driving” and “The driver is drowsy because he is yawning while driving,” and code-formatted Boolean outputs were feasible but less reliable when multiple variables were requested simultaneously. The study also states that current open VLMs are time-consuming and unsuitable for real-time gaze estimation, even though they show some camera-position invariance (Cañas et al., 15 Mar 2025).

4. Clause-grounded industrial safety and online anomaly scoring

In mining safety, "MonitorVLM: A Vision Language Framework for Safety Violation Detection in Mining Operations" defines MonitorVLM as a surveillance-oriented, clause-grounded violation detection system that maps video evidence to the exact mining regulations breached (Wu et al., 4 Oct 2025). The system is built around three components: a LoRA-fine-tuned Qwen2.5-VL-Instruct backbone, a Clause Filter (CF), and a Behavior Magnifier (BM). Its domain-specific dataset contains 9,000 vision–question–answer (VQA) samples across 40 high-frequency mining regulations, derived from Dataset I plus augmentation and auxiliary detection cues. The augmentations are horizontal flipping, low-light synthesis (brightness reduced by 20–50%), and mask occlusion (10–30% of non-critical regions). The CF uses a frozen ResNet-50 for images, a frozen BERT for clauses, and a five-layer MLP (2816 → 1024 → 512 → 256 → 1) to select Top-$T+1$0 relevant clauses; the BM uses LLMDet, 2× enlargement, and Real-ESRGAN to enhance worker regions before reinsertion into the original frame. Quantitatively, CF reduces inference latency by 13.56% while maintaining accuracy; BM yields additional gains of 3.45% in precision and 8.62% in recall; and the final system reaches 93.05% precision, 89.57% recall, and 91.28% F1, outperforming the 72B unfine-tuned baseline by 22.01% in precision, 34.22% in recall, and 28.37% in F1 (Wu et al., 4 Oct 2025).

A methodologically adjacent but differently named framework is MoniTor, introduced for training-free online video anomaly detection (Yang et al., 24 Oct 2025). MoniTor operates on streaming video with BLIP-2 captioning, ImageBind-based caption cleaning, GLM-4-Flash summarization and scoring, an anomaly prior, a dynamic memory gating module, and a standard scoring queue. It computes caption-image similarity as

$T+1$1

retains long-term memory through the thresholded similarity

$T+1$2

and smooths the anomaly score with

$T+1$3

using $T+1$4 and a forgetting threshold $T+1$5. The system reports anomaly scores at 0.6-second decision periods with about 5–6 seconds end-to-end processing latency. On UCF-Crime, it achieves 82.57% AUC; on XD-Violence, it achieves 55.01% AP and 79.11% AUC, outperforming the online training-free baseline online-LAVAD on both datasets (Yang et al., 24 Oct 2025).

5. Monitoring the learner and monitoring the model

In continual learning, "LargeMonitor: Monitoring Online Task-Free Continual Learning via Large Pretrained Models" uses foundation models as an external monitor for online task-free continual learning (TFCL) streams (Yuan et al., 8 Jun 2026). The stream is formalized as

$T+1$6

with minibatches $T+1$7, a frozen large vision model $T+1$8, a FIFO buffer $T+1$9, and a zero-shot drift detector based on CKA and one-sided CUSUM. The similarity statistic is

$$\text{judgeAnomaly}(x, I, \phi(S, C, C_d)) \rightarrow y,$$0

and the CUSUM state is

$$\text{judgeAnomaly}(x, I, \phi(S, C, C_d)) \rightarrow y,$$1

with alarm threshold

$$\text{judgeAnomaly}(x, I, \phi(S, C, C_d)) \rightarrow y,$$2

Upon detection, an LMM such as Qwen-VL, LLaVA, or GPT-4o diagnoses the shift as novel class emergence, domain/environment shift, data corruption, or noise/false alarm, and the learner applies a shift-specific policy. Empirically, buffer size 512 is generally best, 64 severely degrades accuracy, and for gradual domain shifts on CORe50, only the largest ViT-7B/16 attains zero regret. The monitoring signal improves downstream learners such as MVP and MVP-R; on HS-Incremental, MVP-R improves from A_AUC 80.51±0.11 and A_Last 82.06±0.15 to 82.14±0.24 and 82.24±0.26 when augmented with LargeMonitor (Yuan et al., 8 Jun 2026).

A distinct meta-monitoring direction appears in "VLM's Eye Examination: Instruct and Inspect Visual Competency of Vision LLMs," where the monitored object is the VLM’s own perceptual competence (Hyeon-Woo et al., 2024). The framework introduces LENS (Learning ElemeNt for visual Sensory), with Color, Shape, and Semantic tasks, plus a three-stage process of Instruction, Readiness check, and Examination. The readiness stage uses LoRA fine-tuning and accuracy on held-out LENS sets; the examination stage defines Sensitivity Area of Color (SAC) and Sensitivity Area of Shape (SAS) and produces patch-wise semantic score maps. The paper reports a consistent insensitivity to green across models, reflected in larger SAC values for green than for red or blue; for example, LLaVA-13B has Red 0.0036, Green 0.0202, Blue 0.0105, and InstructBLIP-7B has Red 0.0224, Green 0.0793, Blue 0.0336. It also shows that larger LLM backbones, even with the same visual encoder, are more shape-sensitive and semantically discriminative (Hyeon-Woo et al., 2024). In this sense, MonitorVLM becomes an auditing apparatus for multimodal perception itself.

6. Embodied robotic monitoring and recovery

In horticultural robotics, "Visual-Language-Guided Task Planning for Horticultural Robots" provides a modular framework that can be read as a MonitorVLM instantiation for crop monitoring tasks (Cuaran et al., 17 Jan 2026). The system combines two RGB-D cameras, Detic, Kalman-filtered object tracking, OctoMap, and a top-down 2D semantic occupancy map, then exposes a set of tools such as navigate-to-map-point, rotate-and-move-forward, tip camera motions, and a centering tool to a GPT-4.1 agent with tool-calling. The benchmark contains 198 tasks across three Gazebo greenhouse simulations and four categories ranging from Single Plant, Single Target to Multiple Plants, Multiple Targets. Overall, humans achieve SR 90.37%, TCR 97.14±11.67, and SPL 89.97, whereas the zero-shot VLM achieves SR 51.64%, TCR 86.57±16.09, and SPL 35.74. The short-horizon case is substantially stronger: for Single Plant, Single Target, the zero-shot VLM reaches SR 86.67% against 91.94% for humans. Performance collapses in the hardest long-horizon case: for Multiple Plants, Multiple Targets, the zero-shot VLM drops to SR 9.38% against 84.85% for humans. The paper further reports that, under noisy maps in complex polyculture, front-camera polar actions outperform a robot-centric map local navigation mode by about 7% average SR/TCR across categories (Cuaran et al., 17 Jan 2026).

In "VLM Can Be a Good Assistant: Enhancing Embodied Visual Tracking with Self-Improving Vision-LLMs," the monitored process is continuous target tracking in dynamic 3D scenes, and the VLM is invoked only when failure is detected (Wu et al., 27 May 2025). The base tracker is either an offline RL CNN-LSTM actor-critic or a PID controller, both driven by a shared segmentation pipeline. Failure is declared when the target mask is invisible for more than 3 consecutive steps. The VLM then receives three historical observations sampled at five-step intervals before failure, performs a structured failure analysis

$$\text{judgeAnomaly}(x, I, \phi(S, C, C_d)) \rightarrow y,$$3

generates a movement plan $$\text{judgeAnomaly}(x, I, \phi(S, C, C_d)) \rightarrow y,$$4, retrieves the top-3 similar memory items

$$\text{judgeAnomaly}(x, I, \phi(S, C, C_d)) \rightarrow y,$$5

using TF-IDF cosine similarity, and refines a fixed-length recovery action sequence. Across four Unreal Engine 5 environments, the full system improves success rate by up to 72% over the SOTA RL baseline and by up to 220% over the PID baseline. With the RL tracker, success rate rises from 0.36 to 0.62 in the Chemical Plant and from 0.60 to 0.94 in the Supermarket; recovery success rates range from 52.2% to 71.4% depending on environment (Wu et al., 27 May 2025).

7. Recurrent architectural motifs and open limitations

Across these usages, several architectural motifs recur. One is decoupled monitoring. The avionics MonitorVLM runs in the ASM rather than inside certified application partitions; LargeMonitor detects drift in a frozen LVM embedding space rather than through learner-internal loss signals; the EVT assistant calls the VLM only after failure detection; and MoniTor uses separate captioning, memory, and scoring stages rather than a monolithic end-to-end detector (Aloui et al., 2017, Yuan et al., 8 Jun 2026, Wu et al., 27 May 2025, Yang et al., 24 Oct 2025). This suggests that MonitorVLM is often designed as an auxiliary decision layer placed at an observation point that is informative yet minimally disruptive.

A second motif is explicit context construction. In the laboratory setting, anomaly decisions improve monotonically as prompts accumulate Experiment Context, Stage Description, Detection Content, and Anomaly Label Description. In mining, the CF restricts reasoning to the Top-$$\text{judgeAnomaly}(x, I, \phi(S, C, C_d)) \rightarrow y,$$6 most relevant clauses, while BM magnifies worker-centered evidence. In horticultural robotics, performance depends heavily on the quality of the semantic occupancy map and on concise tool context. In MoniTor, long-term and short-term memories, anomaly priors, and a scoring queue provide a textual surrogate for temporal state (Lin et al., 4 Jun 2025, Wu et al., 4 Oct 2025, Cuaran et al., 17 Jan 2026, Yang et al., 24 Oct 2025).

A third motif is that monitoring quality is limited by grounding, latency, and coverage. The avionics paper emphasizes determinism but does not provide analytical latency bounds, computational complexity, or quantitative overhead. The driver-monitoring study reports that current open VLMs are time-consuming and inconsistent on fine-grained gaze. The horticultural system degrades sharply on long-horizon tasks and on noisy semantic maps. MoniTor still incurs about 5–6 seconds end-to-end latency. The mining framework is restricted to 40 high-frequency clauses and acknowledges that stronger temporal reasoning is future work (Aloui et al., 2017, Cañas et al., 15 Mar 2025, Cuaran et al., 17 Jan 2026, Yang et al., 24 Oct 2025, Wu et al., 4 Oct 2025).

A final cross-cutting observation concerns memory and self-correction. LargeMonitor resets baseline statistics after alarms and maps diagnoses to shift-specific adaptation policies; the EVT assistant stores recovery episodes and reflection insights; MoniTor maintains both a forgetting-gated memory and a score-conditioned queue; the VLM eye-examination framework introduces readiness gating before evaluation (Yuan et al., 8 Jun 2026, Wu et al., 27 May 2025, Yang et al., 24 Oct 2025, Hyeon-Woo et al., 2024). This suggests that, across domains, MonitorVLM is less a single model family than a recurring research pattern: a monitoring subsystem that combines privileged observations, structured context, selective multimodal reasoning, and explicit memory to detect inconsistency, anomaly, drift, or safety violations in real time or near-real time.