Blink: Multifaceted Phenomenon in Science

• Blink is a multifaceted phenomenon defined by rapid state transitions, observed in EEG artifacts, dynamic image detection, and multimodal AI benchmarks.
• In physiological and behavioral studies, blink detection uses geometric descriptors and deep learning to enhance brain decoding and driver monitoring accuracy.
• Across domains like distributed systems, astronomy, and quantum dot studies, blink frameworks optimize performance and reveal complex system dynamics.

“Blink” is a cross-disciplinary term whose meaning depends strongly on context. In the literature surveyed here, it denotes a physiological eyelid closure and its electrophysiological correlates, a rapid alternation procedure for detecting change in image sequences, and a recurrent name or acronym for technical frameworks in multimodal modeling, distributed systems, FPGA instrumentation, large-scale inference, and cellular dynamics. The shared semantic core is rapid state transition, rapid comparison, or dynamic allocation of attention or resources, but the underlying objects range from ocular artifacts in EEG to Byzantine consensus and perovskite quantum-dot photophysics (Uppal et al., 19 Aug 2025, Copandean et al., 2019, Fu et al., 2024, Monti et al., 2024, Galimberti et al., 2024, Siavashi et al., 8 Apr 2026, Wang et al., 2019, Olejniczak et al., 11 Feb 2026).

Domain	Meaning of “Blink”	Representative source
EEG and cognition	Spontaneous blinks or blink artifacts as signals	(Uppal et al., 19 Aug 2025)
Vision and behavior	Detected eyelid motion for drowsiness, communication, or welfare	(Wolter et al., 24 Nov 2025)
Astronomy	Alternating registered telescope frames to reveal moving objects	(Copandean et al., 2019)
Multimodal AI	Benchmark or dynamic visual-token mechanism for perception	(Fu et al., 2024)
Distributed systems	Consensus, collectives, scheduling, and serving frameworks	(Monti et al., 2024)
Materials science	Photoluminescence intermittency in quantum dots	(Olejniczak et al., 11 Feb 2026)

1. Blink as an EEG event, artifact, and cognitive marker

In scalp EEG, eyeblinks appear as large-amplitude, short-duration deflections, often tens to hundreds of $\mu\mathrm{V}$, primarily in frontal channels. A recent reanalysis of musicians reading Bach chorales while either listening to or imagining the music treated blink times $\{t_i\}$ as spike trains and compared trials with the Victor–Purpura distance $D_q$, in which insertion or deletion costs 1 and shifting a blink by $\Delta t$ costs $q\Delta t$. The study used standard preprocessing with high-pass at $0.1\,\mathrm{Hz}$, low-pass at $30\,\mathrm{Hz}$, downsampling to $64\,\mathrm{Hz}$, ICA, ICLabel, BLINKER, and manual correction. In 21 expert musicians performing $11 \times 4 \times 2 = 88$ trials per subject, one-trial-left-out nearest-neighbor decoding identified which of four chorales was being read at above-chance accuracy; among the six pilot-analyzed subjects who blinked consistently, mean accuracy was $\approx 33.5\%$ in listening and $\{t_i\}$0 in imagery, versus $\{t_i\}$1 chance, with best-subject accuracies of $\{t_i\}$2 and $\{t_i\}$3, respectively. For 5 of 6 subjects, peak accuracy occurred at $\{t_i\}$4, indicating that blink counts alone often dominated temporal precision (Uppal et al., 19 Aug 2025).

This line of work formalizes a broader claim that spontaneous blinks are not merely nuisance artifacts. The same study explicitly situates blink timing within a literature showing modulation by attention and cognition, including blink-rate suppression during anticipation of salient stimuli, during high cognitive load, and during preparation of motor responses. Its practical implication is that blink trains may contribute to brain decoding, especially in wearables where ocular signals can be easier to record than low-amplitude cortical EEG (Uppal et al., 19 Aug 2025).

A complementary tradition treats blinks as contamination to be removed. The Algorithm for Blink Correction (ABC) corrects EEG directly in the time domain without EOG electrodes by detecting blinks on FP1, computing peak-to-peak amplitude

$\{t_i\}$5

binning amplitudes in $\{t_i\}$6-wide intervals, constructing subject-specific grand-average templates $\{t_i\}$7, and subtracting the appropriate template from all channels. In a three-subject motor-imagery setting classified with Minimum Distance to Riemannian Mean (MDRM), this preprocessing yielded a mean classification accuracy increase of $\{t_i\}$8 (Guttmann-Flury et al., 2019).

Deep learning work has also reframed blink detection in EEG as sequence-to-sequence segmentation rather than artifact rejection. On a UCSD dataset with 31 subjects, 15 healthy controls, and 16 Parkinson’s disease patients, models were trained on raw frontal EEG with 1, 3, or 5 channels. A CNN–RNN hybrid achieved the best mean F1-micro of $\{t_i\}$9, with blink-detection accuracies of $D_q$0, $D_q$1, and $D_q$2 for healthy participants and $D_q$3, $D_q$4, and $D_q$5 for Parkinson’s disease patients in the 1-, 3-, and 5-channel settings, respectively. The same study reported blink rates of $D_q$6 blinks/min in healthy participants and $D_q$7 blinks/min in Parkinson’s disease, with coefficient of variation of inter-blink intervals $D_q$8 versus $D_q$9 (Lensky et al., 5 Sep 2025).

2. Blink detection in human and animal monitoring

In driver monitoring, “blink” is often operationalized through geometric descriptors of eyelid motion. The Eyelid Angle (ELA) metric derives eye openness from 3D facial landmarks obtained with MediaPipe Face Mesh V2. For each upper and lower eyelid, a best-fit plane is estimated by SVD, yielding normals $\Delta t$0 and $\Delta t$1, and the raw eyelid angle is

$\Delta t$2

To combine the two eyes under head rotation, the method weights left and right ELA values by sigmoid functions of the yaw angle $\Delta t$3. Compared with the 2D Eye Aspect Ratio (EAR), ELA was reported to have lower variance under viewpoint changes; in synthetic head sweeps with fixed true eyelid angle $\Delta t$4, ELA yielded vertical-sweep MAE $\Delta t$5 and horizontal-sweep MAE $\Delta t$6, while EAR variance was roughly five times higher. Using ELA-derived temporal features such as closing duration, closed duration, reopening duration, PERCLOS, and PEROPENING, blink detection on the Driver Monitoring Dataset reached average detection accuracy $\Delta t$7, and binary alert-versus-drowsy classification on UTA-RLDD reached $\Delta t$8 (Wolter et al., 24 Nov 2025).

That work also links blink analysis to synthetic-data generation. ELA trajectories were used to animate rigged Blender avatars with controllable blink durations, inter-blink intervals, camera perturbations, and noise. On synthetic videos rendered at $\Delta t$9, $q\Delta t$0, and $q\Delta t$1 Hz, training and testing at the same frame rate produced high synthetic performance, including $q\Delta t$2 alert/drowsy accuracy and $q\Delta t$3 blink detection accuracy at $q\Delta t$4 Hz. The same results also showed substantial frame-rate sensitivity in mixed-FPS settings, which the paper identifies as a limitation of temporally derived blink features (Wolter et al., 24 Nov 2025).

Event-based vision offers a different definition of blink detection: asynchronous sensing of eyelid-motion edges rather than frame-wise appearance classification. In a driver-monitoring pipeline built on event cameras, a GR-YOLO architecture processes windows of $q\Delta t$5 recent events accumulated on a $q\Delta t$6 grid, localizes faces and eyes, and then analyzes eye-region polarity statistics every $q\Delta t$7. Blink candidates are detected from polarity-specific activity $q\Delta t$8 and filtered by vertical spread $q\Delta t$9. On driving data from three subjects, blink detection reached precision $0.1\,\mathrm{Hz}$0, recall $0.1\,\mathrm{Hz}$1, effective temporal resolution $0.1\,\mathrm{Hz}$2, and end-to-end latency $0.1\,\mathrm{Hz}$3 (Ryan et al., 2020).

Automated blink detection has also been extended beyond humans. In equine affective-state assessment, half- and full-blinks are treated as subtle EquiFACS-relevant micro-expressions. Three methods were evaluated on a horse-video dataset recorded at $0.1\,\mathrm{Hz}$4: a frame-based YOLOv12 detector, an optical-flow thresholding baseline, and a fine-tuned VideoMAE model. Final test-set results on 12 public videos gave macro-F1 $0.1\,\mathrm{Hz}$5 for three-class blink classification and $0.1\,\mathrm{Hz}$6 for binary blink detection, with the three classes defined as none, half-blink, and full-blink (Alves et al., 3 Jun 2026).

3. Blink sequences as communication and collective-attention signals

In assistive communication for amyotrophic lateral sclerosis, blink is treated as a discrete control primitive. The BWCNN system classifies each $0.1\,\mathrm{Hz}$7 grayscale eye crop as open or closed at $0.1\,\mathrm{Hz}$8, compresses repeated states, and maps blink counts to a small vocabulary. Session start and end are defined by holding the eyes closed for at least $0.1\,\mathrm{Hz}$9, and an inter-word gap is defined by at least $30\,\mathrm{Hz}$0 of open eyes. The reported dictionary includes 1 blink $30\,\mathrm{Hz}$1 “Yes,” 2 blinks $30\,\mathrm{Hz}$2 “No,” and 3 blinks $30\,\mathrm{Hz}$3 “Hi.” Among several CNN architectures, InceptionV3 gave the best accuracy/latency trade-off with $30\,\mathrm{Hz}$4 accuracy and $30\,\mathrm{Hz}$5 average latency per frame (Ramli et al., 2020).

A very different use of blink exploits its collective suppression and synchronization as a proxy for audience attention. In figure-skating videos, a 1D-CNN was trained on spatio-temporal pose features extracted from $30\,\mathrm{Hz}$6 joints over $30\,\mathrm{Hz}$7 frames to estimate frame-wise blink probability $30\,\mathrm{Hz}$8. Leave-one-video-out experiments showed a significant positive correlation between estimated and actual blink rate in 45 of 48 clips ($30\,\mathrm{Hz}$9), and in $64\,\mathrm{Hz}$0 of 272 manually labeled jump events. Highlights were then defined as stretches of at least five consecutive frames for which $64\,\mathrm{Hz}$1. The method detected 340 scenes in 48 videos, and human evaluation classified $64\,\mathrm{Hz}$2 of these as athletic actions, $64\,\mathrm{Hz}$3 as distinctive choreography, and $64\,\mathrm{Hz}$4 as non-events (Nakano et al., 2020).

These two literatures use the same elementary event—eye closure—but impose very different semantics on it. In one case, the blink sequence is an intentional symbolic code; in the other, aggregated blink suppression is an involuntary indicator of momentary human interest. This suggests that the informativeness of blink lies less in the isolated closure than in its temporal organization.

4. Blink as a comparative method in astronomy

In asteroid detection, the classic “blink” method refers to rapid alternation between registered telescope frames so that fixed celestial sources remain stationary while moving objects appear to jump. The manual workflow requires a sequence of $64\,\mathrm{Hz}$5 exposures of the same field, typically separated by $64\,\mathrm{Hz}$6–$64\,\mathrm{Hz}$7, followed by CCD calibration, distortion correction, registration, and visual blinking on a display. A genuine asteroid appears as a jumping point of light tracing a short straight line across views, after which pixel coordinates are converted to $64\,\mathrm{Hz}$8 and motion is estimated via $64\,\mathrm{Hz}$9 (Copandean et al., 2019).

An automated implementation retains this logic but moves from visual alternation to catalog subtraction and track linking. The pipeline described by Copândean et al. (2017) uses Python wrappers to IRAF, SExtractor, SCAMP, MISSFITS, and SWarp for calibration and field correction, then a Java module, CrossObj, for candidate extraction and linking. Sources matched to a fixed-source reference catalog within $11 \times 4 \times 2 = 88$0 are removed; remaining detections are filtered by $11 \times 4 \times 2 = 88$1 and $11 \times 4 \times 2 = 88$2. Candidate trajectories are linked across five exposures using the constraints $11 \times 4 \times 2 = 88$3, $11 \times 4 \times 2 = 88$4, $11 \times 4 \times 2 = 88$5, and $11 \times 4 \times 2 = 88$6 (Copandean et al., 2019).

The reported performance illustrates the practical value of automation. Initial unfiltered runs took $11 \times 4 \times 2 = 88$7–$11 \times 4 \times 2 = 88$8 and yielded thousands of false trajectories. After imposing $11 \times 4 \times 2 = 88$9, runtime dropped to $\approx 33.5\%$0 with fewer than 20 false positives. With $\approx 33.5\%$1 and $\approx 33.5\%$2, the reported outcome was $\approx 33.5\%$3 correct. The same summary notes that a typical star-rich INT field may contain 30–40 asteroids and that manual blinking by EURONEAR volunteers had yielded 9 real NEA discoveries from La Palma to date (Copandean et al., 2019).

5. BLINK in multimodal perception research

In multimodal machine learning, BLINK names both an evaluation benchmark and a model-side enhancement mechanism, but in both cases the central concern is low-level visual perception rather than language-mediated inference. The benchmark BLINK reformulates 14 classical computer-vision tasks into 3,807 multiple-choice questions over 7,358 unique images, covering visual correspondence, semantic and functional correspondence, relative reflectance, jigsaw completion, art-style grouping, relative depth estimation, object localization, counting, spatial relations, multi-view reasoning, visual similarity, forensics detection, and IQ-style pattern completion. Human accuracy is $\approx 33.5\%$4, random guessing averages $\approx 33.5\%$5, and current multimodal LLMs perform far below the human level: GPT-4V(ision) reaches $\approx 33.5\%$6, Gemini Pro $\approx 33.5\%$7, and Claude 3 Opus $\approx 33.5\%$8. The benchmark’s analysis also reports that specialist models outperform GPT-4V by $\approx 33.5\%$9 percentage points on relative depth, $\{t_i\}$00 points on visual correspondence, and $\{t_i\}$01 points on multi-view reasoning (Fu et al., 2024).

The benchmark is designed to isolate perceptual tasks that humans can often solve “within a blink,” and its error analysis underscores that the deficiency is not reducible to generic reasoning failure. Among 140 sampled GPT-4V failures, the largest categories were hallucinated fine-grained patterns ($\{t_i\}$02), mislocalization of visual prompts ($\{t_i\}$03), spatial-relation confusions ($\{t_i\}$04), reasoning breakdown despite correct image parsing ($\{t_i\}$05), and failure to ground question referents ($\{t_i\}$06) (Fu et al., 2024).

A later framework called Blink proposes to improve such perception deficits through dynamic visual token resolution within a single forward pass. It computes token-level saliency from attention maps, aggregates token saliency into patch saliency, and defines a saliency ratio

$\{t_i\}$07

If $\{t_i\}$08 exceeds an expansion threshold $\{t_i\}$09, the most salient patch is upsampled and processed by a TokenSR module; if $\{t_i\}$10 falls below $\{t_i\}$11, previously inserted high-resolution tokens are dropped. Evaluated with LLaVA-1.5-7B on seven benchmarks, the method improved MME$\{t_i\}$12 from $\{t_i\}$13 to $\{t_i\}$14, MME$\{t_i\}$15 from $\{t_i\}$16 to $\{t_i\}$17, ScienceQA from $\{t_i\}$18 to $\{t_i\}$19, and MM-Vet from $\{t_i\}$20 to $\{t_i\}$21, while a training-free interpolation baseline did not match these gains (Feng et al., 11 Dec 2025).

Together, these works use “BLINK” to mark a gap between recognition and genuine perception, and then to propose an architectural mechanism for adaptive fine-grained inspection. This suggests a second modern sense of blink: not eye closure, but fast alternation between global scanning and local focus.

6. Blink as a family of systems and optimization frameworks

Several unrelated systems papers use “Blink” as the name of an optimization framework. Their commonality is procedural rather than semantic: each Blink implements a fast path, a lightweight sampling strategy, or a topology-aware allocation mechanism.

Area	Core mechanism	Representative result
Byzantine consensus	Leaderless, signature-free representative binary consensus	Decision in $\{t_i\}$22 in the good case (Monti et al., 2024)
FPGA power monitoring	Behavioral simulation + direct FPGA measurement	Average $\{t_i\}$23 time-to-solution speedup (Galimberti et al., 2024)
Big-data cost optimization	Three tiny sample runs and linear memory models	Optimal cluster size in 15 of 16 cases (Al-Sayeh et al., 2022)
Distributed ML collectives	Packing spanning trees over heterogeneous links	Up to $\{t_i\}$24 faster synchronization (Wang et al., 2019)
CPU-free LLM serving	SmartNIC offload + persistent GPU scheduler	P99 TTFT reduced by up to $\{t_i\}$25 (Siavashi et al., 8 Apr 2026)

In Byzantine fault tolerance, Blink is the representative binary consensus primitive underlying Flutter. It assumes partial synchrony, $\{t_i\}$26 servers, authenticated FIFO links without signatures, and an adaptive adversary. The fast path decides immediately when a server has received $\{t_i\}$27 matching $\{t_i\}$28 messages; otherwise, after collecting $\{t_i\}$29 total suggestions, the server proposes the majority value into an inner weak-validity consensus $\{t_i\}$30. In the good case, where all correct servers propose the same value and messages take exactly $\{t_i\}$31, termination occurs in one network delay $\{t_i\}$32. The paper presents this as a generalization of Bosco from $\{t_i\}$33 to $\{t_i\}$34 servers (Monti et al., 2024).

In FPGA-based power monitoring, Blink is a five-phase automated design flow comprising target implementation, behavioral simulation, FPGA power measurement, model identification, and monitor implementation. It replaces gate-level simulation and VCD-to-power-trace extraction with behavioral simulation and direct FPGA measurement, using sparse linear regression to construct an estimator

$\{t_i\}$35

Across ten FPGA designs, reported RMSE was $\{t_i\}$36, LUT overhead $\{t_i\}$37, flip-flop overhead $\{t_i\}$38, dynamic-power overhead $\{t_i\}$39, and overall time-to-solution speedup averaged $\{t_i\}$40, ranging from $\{t_i\}$41 to $\{t_i\}$42 (Galimberti et al., 2024).

In cluster-size optimization for iterative big-data applications, Blink predicts cached-dataset size and execution-memory footprint from three single-node sample runs using $\{t_i\}$43–$\{t_i\}$44 of the data. The framework then selects the smallest cluster size that yields an eviction-free run. On 16 real-scale experiments, it selected the optimal cluster size in 15 cases, with average sample-run cost $\{t_i\}$45 of the cost of the optimal run and savings of up to $\{t_i\}$46 relative to average execution cost (Al-Sayeh et al., 2022).

In distributed training, Blink is a collective-communication library that dynamically generates communication primitives by packing spanning trees over heterogeneous GPU interconnects. Compared to NCCL, the reported gains reach up to $\{t_i\}$47 faster model synchronization and up to $\{t_i\}$48 reduction in end-to-end training time for image-classification workloads (Wang et al., 2019).

In LLM serving, Blink is an end-to-end architecture that removes the host CPU from the steady-state inference path. Request handling is moved to a SmartNIC DPU, while batching, scheduling, CUDA-graph launch, and KV-cache management are handled by a persistent GPU kernel. Against TensorRT-LLM, vLLM, and SGLang, the reported isolated gains include P99 TTFT reduced by up to $\{t_i\}$49, P99 TPOT reduced by up to $\{t_i\}$50, decode throughput improved by up to $\{t_i\}$51, and energy per token reduced by up to $\{t_i\}$52. Under CPU interference, Blink is reported to remain close to isolated performance, while baselines can degrade by up to two orders of magnitude (Siavashi et al., 8 Apr 2026).

7. Blinking in photophysics and latent biological dynamics

In colloidal quantum dots, blinking refers to random switching of photoluminescence between bright and dark levels. For lead-halide perovskite quantum dots, conventional A-type and BC-type models both predict that dark states should exhibit worse antibunching, that is, higher zero-delay correlation $\{t_i\}$53. A recent state-resolved spectroscopy study identified a different regime in individual Ni-doped $\{t_i\}$54 PQDs. In $\{t_i\}$55 of dots, fluorescence lifetime–intensity distribution maps were consistent with conventional A/BC blinking and $\{t_i\}$56 increased in dark or charged states. In the remaining $\{t_i\}$57, denoted “single-photon blinking,” dark states became both dimmer and cleaner single-photon emitters: a representative dot showed $\{t_i\}$58 decreasing from $\{t_i\}$59 in the bright state to $\{t_i\}$60 in the dark state, while exciton quantum yield dropped by a factor of $\{t_i\}$61 and biexciton quantum yield by a factor of $\{t_i\}$62 (Olejniczak et al., 11 Feb 2026).

The proposed explanation is a self-trapped-exciton (STE) mechanism that diverts hot excitons into long-lived, weakly emissive configurations and suppresses biexciton formation. In the model,

$\{t_i\}$63

and using the reported bright- and dark-state biexciton quantum yields yields $\{t_i\}$64. The paper interprets this as a lattice-driven route to intrinsically suppressed multiphoton emission (Olejniczak et al., 11 Feb 2026).

In cell biology, BLINK names a trajectory-based recurrent state-space model for natural-killer-cell cytotoxicity rather than a literal blinking event. The model combines a DreamerV2-style recurrent state-space world model with a nonnegative apoptosis increment head $\{t_i\}$65, so that cumulative death is built by monotonic accumulation. On long-term NK–tumor recordings comprising 485 train, 29 validation, and 57 test tracks, BLINK achieved test MAE $\{t_i\}$66, RMSE $\{t_i\}$67, Pearson correlation $\{t_i\}$68, within-$\{t_i\}$69-event accuracy $\{t_i\}$70, and forecast error $\{t_i\}$71. Latent-space analysis produced four behavioral modes: High Cytotoxic, Motile, Low Cytotoxic, and Quiescent (Nematollahi et al., 5 Mar 2026).

Across these two literatures, “blinking” again denotes intermittency, but the intermittency is no longer ocular. In perovskite quantum dots it is optical-state switching; in BLINK for cytotoxicity it is a model name for latent interaction dynamics. The term therefore spans direct physiology, signal-processing convention, and metaphorical extension into dynamical systems.