Drift: Concepts & Applications

• Drift is a multifaceted concept describing systematic deviations from baseline behavior in physical systems, stochastic dynamics, and data-driven models.
• In fluid dynamics and transport physics, drift quantifies effective velocities and informs model closures in gas-solid flows and oscillatory systems.
• In sensing and machine learning, drift captures temporal changes in data distributions and predictions, driving methods for calibration and non-IID learning.

In contemporary research usage, drift denotes several technically distinct but structurally related phenomena: an effective transport velocity in multiphase flow and oscillatory media, a deterministic component in stochastic dynamics, a time-varying deviation in sensing and data-generating processes, a prediction-space discrepancy in federated learning, and the name of multiple benchmarks and model architectures whose titles expand to DRIFT. Across these settings, the term typically marks a departure from a baseline—resolved flow, stationary dynamics, nominal calibration, IID training, or discrete task segmentation—and the central problem is to model, estimate, regularize, or exploit that departure with sufficient fidelity for downstream prediction or control (1808.04489, Cholaquidis et al., 2016, Hurst et al., 10 Jun 2025, Sun et al., 13 May 2026).

1. Taxonomy of meanings

The research literature uses drift in at least five recurrent senses.

Domain	Meaning of drift	Representative source
Gas-solid fluidization	Sub-grid velocity quantity tied to filtered Eulerian drag	(1808.04489)
Stochastic processes	Deterministic trend term in reflected diffusion or evidence accumulation	(Cholaquidis et al., 2016, Liu et al., 11 Dec 2025)
Transport physics	Effective mean motion generated by oscillations, collisions, or waves	(Vladimirov, 2010, Ochs et al., 2017, Calvert et al., 2021)
Data and sensing systems	Time-varying change in response function, distribution, or intent state	(Hurst et al., 10 Jun 2025, Hinder et al., 2020, Hossain et al., 14 Feb 2026)
Named methods and benchmarks	Acronym for specific architectures, planners, and evaluation suites	(Sun et al., 13 May 2026, Liu et al., 4 Jun 2026, Pei et al., 10 Mar 2026, Luo et al., 29 Jun 2026)

In the gas-solid fluidization literature, drift velocity is a filtered two-fluid quantity defined by

$$\varphi_s v_{d,i} = \overline{\varphi_s u_{g,i} - \varphi_s \tilde{u}_{g,i}}$$

and equivalently

$\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$

making it a filtered measure of gas-phase velocity fluctuation relative to resolved gas motion (1808.04489).

In stochastic-process language, drift is the term $\mu(X_s)\,ds$ in a reflected Brownian motion, or the latent evidence-accumulation rate in a drift diffusion model. In both cases it is a parameter of the dynamics rather than a post hoc descriptor of error (Cholaquidis et al., 2016, Liu et al., 11 Dec 2025).

In machine learning and monitoring systems, drift usually refers to non-stationarity over time. The relevant object may be the data distribution, a subset of features, the sensor response coefficients, the network’s state relative to an intent, or the discrepancy between local and global predictions in federated learning (Hinder et al., 2020, Hurst et al., 10 Jun 2025, Kim et al., 2023, Hossain et al., 14 Feb 2026).

2. Drift as effective transport in physical systems

In filtered two-fluid modeling for gas-solid fluidization, drift velocity is directly relevant to the filtered Eulerian drag force, which the paper describes as the most important sub-grid constitutive term in fTFM. Starting from the standard gas-phase continuity and momentum equations, the drift-velocity transport equation is derived without additional assumptions by forming a transport equation for the volume average of the Favre fluctuation of gas velocity and rewriting it in conservative form. The resulting equation contains transient and convection terms on the left-hand side and seven labeled contributions on the right-hand side: large-scale shear production, large-scale volume-fraction-gradient production, micro-scale volume-fraction-fluctuation production, sub-grid drag source, velocity-fluctuation/dilatation correlation, turbulent diffusion, and stress-fluctuation diffusion. Only the main shear-production term is exact; the other six require closure, with term $(a3)$ emphasized as unique to gas-solid flow and lacking a direct single-phase analog (1808.04489).

In high-frequency oscillatory fluid flows, drift is the emergent averaged transport velocity appearing after two-timing and Eulerian averaging. The asymptotic structure depends on the path taken in the $(1/\omega,\delta)$-plane, and the paper identifies critical, super-critical, and sub-critical asymptotic families. At least three distinct drift velocities,

$$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$

can arise as leading $O(1)$ transport depending on the scaling. In the critical family, the averaged motion is a pure drift in the zeroth and first approximations and drift combined with pseudo-diffusion in the second approximation; in the super-critical families calculated in the paper, the leading behavior is pure drift without pseudo-diffusion (Vladimirov, 2010).

In magnetized plasmas with collisionality gradients, drift appears as a single-particle gyrocenter drift generated by spatially asymmetric collisional drag and background flow. In the heuristic low-$\nu_0/\Omega$ limit, the drift is

$$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$

The first term is a gradient-drift up the collisionality gradient; the second is a background-flow contribution. In the high-$Z_i$ regime this reduces to the impurity pinch, while in the low-temperature singly-ionized regime the drift-to-diffusion asymmetry becomes mass-dependent and energy-dependent, enabling a proposed mass-separation scheme (Ochs et al., 2017).

In crowded random media, a constant drift does not simply add persistent ballistic motion. For overdamped tracer motion through concave fibrinogen-like obstacles, stronger drift can increase the likelihood of trapping because more attempted steps intersect obstacles and are rejected. The resulting dynamics are anisotropic, exhibit large trajectory-to-trajectory variability, and can show superdiffusive and subdiffusive signatures in the same system depending on whether one examines ensemble MSD, TAMSD, detrended variance, or individual trajectories (Kubala et al., 2020).

For floating marine litter in deep-water waves, wave-induced drift can exceed classical Stokes drift. The mechanism is the combination of variable submergence and a dynamic buoyancy force resolved normal to the sloping free surface, which gives a nonzero mean horizontal component when the submergence response is phase-lagged. The paper derives a closed-form approximation for the second-order mean drift and reports that, for a wave with period about $\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$0 s and steepness $\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$1, a $\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$2 m diameter object with density $\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$3 can have about a $\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$4 increase in wave-induced drift relative to a tracer, whereas a $\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$5 m object may behave like a tracer (Calvert et al., 2021).

3. Stochastic drift, estimation, and inferential failure modes

For reflected Brownian motion with drift on a bounded $\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$6 domain, the process is defined by the Skorokhod-type SDE

$\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$7

Here drift is the Lipschitz field $\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$8, while reflection is mediated by the inward normal $\varphi_s v_{d,i} = \widetilde{u'_{g,i}},$9 and the boundary local time $\mu(X_s)\,ds$0. The paper establishes Harris recurrence, a non-trap condition, and geometric ergodicity

$\mu(X_s)\,ds$1

then uses these properties to justify kernel estimation of the stationary density and a local increment-based estimator of drift from a single observed trajectory. Under

$\mu(X_s)\,ds$2

the estimator $\mu(X_s)\,ds$3 is consistent in probability for all interior $\mu(X_s)\,ds$4 (Cholaquidis et al., 2016).

In drift diffusion models, drift is the central latent parameter governing evidence accumulation, but the paper "Time-Averaged Drift Approximations are Inconsistent for Inference in Drift Diffusion Models" shows that replacing a time-varying within-trial drift by its temporal average can yield inconsistent inference. For the one-sided piecewise-constant example, the TADA estimator converges almost surely to a limit $\mu(X_s)\,ds$5 satisfying

$\mu(X_s)\,ds$6

This establishes that TADA does not converge to the true drift. In the attentional DDM numerical example, the TADA-based estimate systematically underestimates $\mu(X_s)\,ds$7, thereby overstating the effect of attention; when $\mu(X_s)\,ds$8 the bias vanishes because the model reduces to the standard DDM (Liu et al., 11 Dec 2025).

These two lines of work treat drift in opposite inferential modes. In reflected diffusion, drift is estimated from path recurrence under explicit reflection geometry. In DDMs with time-varying evidence, the central warning is that temporal averaging of drift can destroy identifiability of the true dynamics (Cholaquidis et al., 2016, Liu et al., 11 Dec 2025).

4. Drift as non-stationarity in sensing, data streams, and networked systems

In IoT sensing, drift is a time-varying change in the sensor response function itself. For dissolved oxygen sensors, the paper models measurements as

$\mu(X_s)\,ds$9

and treats drift as temporal change in $(a3)$0 and $(a3)$1 due to aging, fouling, poisoning, membrane degradation, or environmental influence. Separate Gaussian Process Regression models are fit to the coefficients using sparse calibration data and their standard errors, and corrected analyte values are recovered by inverting the response function. Offline drift correction delivers MSE reductions of over $(a3)$2 for one sensor and more than $(a3)$3 on average with the Matérn kernel, while uncertainty-driven calibration scheduling yields a further network-wide MSE improvement of $(a3)$4 on average and $(a3)$5 excluding the hardest $(a3)$6-hour interval cases (Hurst et al., 10 Jun 2025).

In data-stream analysis, concept drift is formalized as temporal change in the underlying distribution $(a3)$7. The paper "Analysis of Drifting Features" distinguishes drift inducing features, whose observed drift cannot be explained by conditioning on other features, from faithfully drifting features, which drift only as a consequence of other variables. By treating time $(a3)$8 as the target variable, the work connects drift analysis to feature relevance theory and proves, under strictly positive density, the equivalences: strongly relevant for predicting $(a3)$9 iff drift inducing, weakly relevant iff faithfully drifting, and irrelevant iff non-drifting. The two proposed practical approaches are statistical DFA, based on conditional independence testing and graph structure, and Relevance Bounds, based on approximating $(1/\omega,\delta)$0 with a Random Forest classifier or regressor (Hinder et al., 2020).

In federated learning under Non-IID data, drift is the prediction discrepancy between a client’s previous local model and the current global model on the same input,

$(1/\omega,\delta)$1

Learning from Drift estimates this discrepancy in normalized logit space and regularizes the local model in the reverse direction through an auxiliary soft target $(1/\omega,\delta)$2. The stated motivation is that constraining classifier outputs is more effective than constraining features or parameters for preventing degradation on Non-IID data (Kim et al., 2023).

In Intent-Based Networking, intent drift is the gradual divergence of operational state from the intended target before overt failure. LEAD-Drift reframes detection as fixed-horizon supervised prediction with labels

$(1/\omega,\delta)$3

smooths the raw risk score with an EMA, and triggers an alert by threshold first-crossing. Reported results are Detection Rate $(1/\omega,\delta)$4, Average Lead Time $(1/\omega,\delta)$5 minutes, and False Positive Rate/day $(1/\omega,\delta)$6. Relative to a distance-based baseline, average lead time improves by $(1/\omega,\delta)$7 minutes $(1/\omega,\delta)$8; relative to a weighted-KPI heuristic, alert noise is reduced by $(1/\omega,\delta)$9 with a $$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$0-minute lead-time trade-off (Hossain et al., 14 Feb 2026).

5. DRIFT as a named benchmark and model family in machine learning

Several papers use DRIFT as an acronym for concrete algorithms or benchmarks rather than a generic concept. In continual graph learning, DRIFT is a benchmark for task-free streams with continuous distribution shifts. The stream distribution is modeled as

$$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$1

with Gaussian scheduling

$$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$2

The benchmark spans hard task switches, boundary-local mixing, global mixing, and continuous Gaussian transitions. Under Gaussian-mixed drift, the best task-free methods remain far below Joint training—for example, on Arxiv-CL the best task-free result is around $$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$3 versus Joint $$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$4—and the paper argues that many current continual graph learning methods rely implicitly on task-boundary information (Sun et al., 13 May 2026).

In vision-language modeling, DRIFT is a residual flow adapter for continuous decoding tasks. A base predictor $$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$5 provides a coarse estimate $$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$6, while a flow-matching refiner models the residual around a Gaussian bridge initialized at $$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$7. This residualization converts global transport into localized residual transport. The method improves visual grounding and robotic control: on Charades-STA, $$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$8 increases from $$\overline{V}_0,\qquad \overline{V}_1,\qquad \overline{V}_2,$$9 to $O(1)$0 and mIoU from $O(1)$1 to $O(1)$2; on the RefCOCO series, the Qwen3-VL baseline average rises from $O(1)$3 to $O(1)$4; with OpenVLA on Libero the average action success rate increases from $O(1)$5 to $O(1)$6 (Liu et al., 4 Jun 2026).

In online self-improvement for LLMs, DRIFT stands for Difficulty Routing Self-Distillation with Rhythm-Gated Exploration and Success Buffer Training. It mixes self-distillation on incorrect rollouts with rhythm-gated GRPO-style reinforcement on correct rollouts, using historical pass rates to partition problems into easy, medium, and hard bins and a success buffer to retain high-quality trajectories. On the average score over five benchmarks, DRIFT reaches $O(1)$7, outperforming GRPO by $O(1)$8 and SDPO by $O(1)$9; on ToolUse it reaches $\nu_0/\Omega$0, improving over GRPO by $\nu_0/\Omega$1 and SDPO by $\nu_0/\Omega$2 (Luo et al., 29 Jun 2026).

6. Drift-aware autonomy, perception, and control

In mobile-robot planning, DRIFT can denote Diffusion-based Rule-Inferred for Trajectories, a conditional diffusion framework for mapless trajectory generation in unstructured environments. Its architecture combines a GNN-based Structured Scene Perception module for global topological consistency with a Graph-Conditioned Time-Aware GRU for target-sensitive recurrent denoising. The reported balance point is centimeter-level imitation fidelity with competitive smoothness: Final Displacement Error $\nu_0/\Omega$3 m, Jerk $\nu_0/\Omega$4, Inference Success Rate $\nu_0/\Omega$5, Predicted Collision Rate $\nu_0/\Omega$6, and latency $\nu_0/\Omega$7 s (Zhao et al., 1 Mar 2026).

In 4D radar perception for automated driving, DRIFT stands for Dual-Representation Inter-Fusion Transformer. It uses a point path for fine-grained local geometry, a pillar path for coarse-grained global context, and multi-stage feature-sharing blocks, with cross-attention producing the best fusion results. On the View-of-Delft dataset, DRIFT achieves $\nu_0/\Omega$8 mAP on the entire area and $\nu_0/\Omega$9 mAP on the driving corridor, improving to $$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$0 and $$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$1 with pre-training. On the internal perciv-scenes-2 dataset it improves object detection from CenterPoint’s mAP $$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$2 and NDS $$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$3 to mAP $$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$4 and NDS $$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$5, while remaining real-time capable with latency $$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$6 ms for add fusion and $$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$7 ms for cross-attention fusion (Pei et al., 10 Mar 2026).

In autonomous driving risk assessment, DRIFT may also refer to Driving Risk Inference via Field Transmission, where risk is represented as a spatiotemporal field $$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$8 evolved by an advection-diffusion-reaction PDE with vehicle, occlusion, and merge-topology source terms. The method introduces field-centric metrics such as Lane-Change Risk Differential, Temporal Anticipation Index, Occlusion Sensitivity Index, and Occlusion Response Latency. Reported behavior-consistency and occlusion-robustness results include LCRD $$\mathbf{v}_d = \frac{1}{2}\rho^2\nabla \nu_s + \frac{\langle \nu_s\rangle}{\Omega}\,\mathbf{v}\times\hat{b} +\mathcal{O}\!\left(\frac{\nu_0}{\Omega}\right).$$9, TAI $Z_i$0 (about $Z_i$1 s), OSI $Z_i$2, ORL $Z_i$3 s, and $Z_i$4Coll. $Z_i$5 (Wang et al., 27 May 2026).

A separate line of work treats drift as localization error in SLAM-based navigation and makes its reduction an explicit planning objective. The proposed pipeline learns drift-minimizing feature regions from LIDAR range images using a directional triplet ranking loss and GradCAM, then steers an MPC planner toward those regions while penalizing acceleration and speed error. In CARLA, the method reports drift reduction of up to $Z_i$6 compared to benchmark approaches, with smaller gains when semantically useful features become sparse (Omama et al., 2022).

Across these systems, drift is not merely an error statistic. It is a design target for representation learning, uncertainty propagation, calibration scheduling, replay allocation, and motion planning. A plausible implication is that the term’s persistent reuse across domains reflects a common technical motif: the need to model structured deviation over time or scale, rather than to treat non-stationarity, hidden transport, or localization error as unmodeled residuals.