Task-Oriented Randomization

• Task-oriented randomization is a design principle that aligns randomness with specific inferential, computational, or control objectives rather than serving as a generic source of noise.
• It is applied across experimental design, adaptive data collection, robotics, and language model pretraining to achieve balance, targeted exploration, and stability.
• Methods like sequential rerandomization, VNSRR, and SPR demonstrate how task-specific randomization improves variance reduction, treatment assignment, and model robustness.

Task-oriented randomization denotes a family of methods in which the randomization mechanism is engineered around a downstream objective rather than used as a generic source of stochasticity. Across the literature, the objective may be covariate balance in randomized experiments, finite-sample validity under adaptive data collection, exact equality of treatment counts in clinical trials, treatment-budget and uniformity constraints in micro-randomized trials, task-relevant exploration for sim-to-real transfer, unsupervised generation of feasible and diverse manipulation tasks, positional robustness in dialog pretraining, or stabilization of black-box outputs through perturbation and aggregation (Zhou et al., 2017, Lu et al., 2023, Nair et al., 2023, Santos-Magalhaes et al., 2015, Meng et al., 3 Jan 2025, Liang et al., 2020, Fang et al., 2022, Gu et al., 2020, Wang et al., 24 Jun 2026). The literature therefore suggests that the common feature is not a single stochastic primitive, but the alignment of randomness with a specified inferential, computational, or control task.

1. Conceptual scope

In experimental design, task-oriented randomization appears as assignment rules tailored to balance criteria or enrollment constraints. In adaptive inference, it appears as weighted resampling that respects path-dependent assignment probabilities. In robotics and sequential decision-making, it appears as exploration or task generation targeted to the variables that matter for downstream performance. In language modeling and black-box stabilization, it appears as randomization of positions or inputs to prevent brittle dependence on incidental structure (Lu et al., 2023, Nair et al., 2023, Liang et al., 2020, Gu et al., 2020, Wang et al., 24 Jun 2026).

A recurring contrast in these works is between generic randomization and randomization that is conditional on structure. The structure may be the feasible assignment set in a randomized experiment, the adaptive policy that generated the data, the stochastic pattern of future “risk times” in a mobile health intervention, the geometry of a manipulation task space, or the absolute-position embedding scheme of GPT-2 (Zhou et al., 2017, Meng et al., 3 Jan 2025, Fang et al., 2022, Gu et al., 2020). In each case, the method preserves stochasticity while changing how randomness is allocated.

This suggests that task-oriented randomization is best understood as a design principle. Under this principle, the randomization rule is chosen to satisfy a scientific, inferential, or operational target: minimize Mahalanobis imbalance, restore validity under non-exchangeability, guarantee exactly equal treatment numbers, approximate a treatment budget online, identify task-relevant physical parameters, generate feasible and diverse tasks, or smooth unstable black-box outputs (Santos-Magalhaes et al., 2015, Lu et al., 2023, Nair et al., 2023, Meng et al., 3 Jan 2025, Liang et al., 2020, Fang et al., 2022, Wang et al., 24 Jun 2026).

2. Controlled assignment and rerandomization in experiments

One line of work uses task-oriented randomization to construct treatment assignments with explicit balance properties. “On Preparing a List of Random treatment Assigns” introduces a threshold-based without-replacement algorithm that generates a random permutation of patient identifiers $1,\dots,N$, assumes $g \mid N$, and then divides the resulting list into $g$ equal groups of size $N/g$. The method is formulated through an assignment list $L'=\{L_1,\dots,L_N\}$ and a survivor list $S=\{1,\dots,N\}$, with thresholds $T_k=k/|S|$ applied to successive $U_i\sim \mathrm{Uniform}(0,1)$. The paper defines a “strong randomised list” as one satisfying three propositions: each patient is equally likely to occupy any position, each patient has equal probability $1/g$ of receiving any treatment, and in the multi-phase extension each group has equal probability $1/g$ of receiving any treatment in any phase (Santos-Magalhaes et al., 2015).

A second line treats rerandomization as an imbalance-control problem. “Sequential rerandomization” studies the case in which units arrive in groups $g \mid N$0 rather than all at once. For each stage $g \mid N$1, a tentative assignment is generated for the new group, previous assignments are held fixed, and the cumulative Mahalanobis distance $g \mid N$2 is computed using all units enrolled so far. Acceptance occurs when $g \mid N$3, with stagewise acceptance probability chosen as $g \mid N$4. The key comparison result is that, for the same expected total number of rerandomizations $g \mid N$5, sequential rerandomization satisfies

$g \mid N$6

and for equal group sizes this becomes $g \mid N$7. Under an additive treatment-effect model, variance reduction is tied to the shrinkage factor $g \mid N$8 through

$g \mid N$9

The paper therefore casts sequential randomization as an adaptive but explicitly randomized alternative to deterministic minimization (Zhou et al., 2017).

A third line formulates rerandomization directly as optimization. “Fast Rerandomization via the BRAIN” states that the paper’s actual method is VNSRR (Variable Neighborhood Searching Rerandomization). Here the design task is to minimize the Mahalanobis distance

$g$0

subject to $g$1 and $g$2. Rather than using acceptance-rejection sampling rerandomization (ARSRR), VNSRR begins from a random feasible assignment, performs pairwise treated/control swaps, accepts swaps that reduce $g$3, and uses a shaking step when no improving move is found. The estimator

$g$4

remains unbiased under the paper’s symmetry, initialization, and update conditions, and rerandomization yields variance reduction relative to complete randomization. The method is extended to sequential, stratified, and cluster randomized experiments, with simulations reporting that VNSRR can sample thousands of balanced assignments in seconds and a real-data example reporting about $g$5 seconds for VNSRR versus about $g$6 minutes for RR/ARSRR and more than half an hour for PSRR (Lu et al., 2023).

Taken together, these works make clear that experimental task-oriented randomization is not simply “more randomization.” It is randomization over a constrained design space, guided by explicit balance objectives, enrollment structure, or computational requirements.

3. Adaptive data collection and online treatment budgets

A distinct use of task-oriented randomization arises when data are collected adaptively and are therefore not exchangeable. “Randomization Tests for Adaptively Collected Data” extends randomization-based inference to contextual bandits, reinforcement learning, and adaptive experimental designs by replacing ordinary permutation logic with a weighted randomization test. If $g$7 is the observed adaptive dataset, $g$8 is a test statistic, and resampled datasets $g$9 are generated under a proposal distribution, the framework assigns weights

$N/g$0

and forms a weighted p-value

$N/g$1

or its Monte Carlo self-inclusion variant. The central claim is that one should not force adaptive data to be exchangeable; one should instead use the known or estimable assignment mechanism $N/g$2 to build a valid weighted reference distribution (Nair et al., 2023).

The same task-specific logic appears in online sampling for micro-randomized trials. “Evaluation of the HeartSteps Online Sampling Algorithm” studies HeartSteps V2V3, in which participants have decision times every five minutes during a 12-hour intervention window, sedentariness is defined as fewer than $N/g$3 steps in the prior $N/g$4 minutes, and randomization occurs only when the participant is both available and sedentary. The algorithm, Sequential Risk Time Sampling (SeqRTS), is designed to satisfy two constraints simultaneously: an average of $N/g$5 interventions per day, equivalently $N/g$6 per four-hour block, and uniform delivery across the decision times at which randomization is allowed. If the number of risk times $N/g$7 in a block were known, the oracle probability would be

$N/g$8

Because $N/g$9 is unknown online, SeqRTS uses

$L'=\{L_1,\dots,L_N\}$0

where $L'=\{L_1,\dots,L_N\}$1 estimates the remaining number of risk times, $L'=\{L_1,\dots,L_N\}$2 is a tuned budget, and $L'=\{L_1,\dots,L_N\}$3. The estimator $L'=\{L_1,\dots,L_N\}$4 uses features including $L'=\{L_1,\dots,L_N\}$5 and the risk run length

$L'=\{L_1,\dots,L_N\}$6

Actual treatment is drawn as $L'=\{L_1,\dots,L_N\}$7, and daily treatment count is $L'=\{L_1,\dots,L_N\}$8 with $L'=\{L_1,\dots,L_N\}$9 (Meng et al., 3 Jan 2025).

The HeartSteps evaluation is notable because it explicitly distinguishes average-treatment control from within-block uniformity. For uniformity the paper argues that KL divergence is confounded by budget mismatch and proposes mean absolute deviation,

$S=\{1,\dots,N\}$0

with $S=\{1,\dots,N\}$1 when all risk-time probabilities within the block are equal. Using GEE with independence and AR-1 working correlations, the estimated mean number of interventions under AR-1 was $S=\{1,\dots,N\}$2 per day, with block means $S=\{1,\dots,N\}$3, $S=\{1,\dots,N\}$4, and $S=\{1,\dots,N\}$5, and the estimated mean MAD values were $S=\{1,\dots,N\}$6 for the whole day and $S=\{1,\dots,N\}$7, $S=\{1,\dots,N\}$8, and $S=\{1,\dots,N\}$9 by block. The paper also reports that the oracle probability can underperform because treatment changes future availability through the one-hour no-notification rule, whereas SeqRTS with tuned $T_k=k/|S|$0 achieves values close to the target (Meng et al., 3 Jan 2025).

These works show that, under adaptive collection, task-oriented randomization is not primarily about symmetry of labels. It is about calibrating the stochastic mechanism to non-exchangeability, uncertain future opportunities, and the estimand of interest.

4. Exploration and task generation in robotics

In sim-to-real transfer, task-oriented randomization takes the form of exploration targeted to task-relevant latent parameters. “Learning Active Task-Oriented Exploration Policies for Bridging the Sim-to-Real Gap” assumes a parameterized dynamics model $T_k=k/|S|$1 with unknown real-world parameters $T_k=k/|S|$2. An exploration policy $T_k=k/|S|$3 is executed first, a system identification map $T_k=k/|S|$4 produces

$T_k=k/|S|$5

and a model-based task policy $T_k=k/|S|$6 is then synthesized and executed. The exploration policy is trained to minimize downstream task regret,

$T_k=k/|S|$7

This directly contrasts with task-agnostic system identification, which seeks generic parameter estimation accuracy rather than identification of the parameters that most affect the final controller (Liang et al., 2020).

The paper analyzes the framework in an LQR setting and instantiates it in pouring and object-dragging tasks. In pouring, the exploration policy decides which cup to lift in order to estimate mass via force sensing; the task-oriented policy pushes the probability of measuring the task-relevant cup to about $T_k=k/|S|$8, whereas the task-agnostic policy settles around $T_k=k/|S|$9. In Franka Panda experiments, the task-oriented policy achieves average task cost of about $U_i\sim \mathrm{Uniform}(0,1)$0 g versus $U_i\sim \mathrm{Uniform}(0,1)$1 g for task-agnostic exploration. In object dragging, real-world evaluation over $U_i\sim \mathrm{Uniform}(0,1)$2 trials shows lower mean and variance of cost for task-oriented exploration, and the learned exploration trajectory starts farther from the center of mass, making box motion more sensitive to the object-table friction parameter (Liang et al., 2020).

A related but distinct use appears in “Active Task Randomization: Learning Robust Skills via Unsupervised Generation of Diverse and Feasible Tasks.” Here the randomization target is not a treatment assignment or a probing trajectory but the training task itself. A task is parameterized as

$U_i\sim \mathrm{Uniform}(0,1)$3

where $U_i\sim \mathrm{Uniform}(0,1)$4 is the object list, $U_i\sim \mathrm{Uniform}(0,1)$5 the spatial relations in the initial scene graph, $U_i\sim \mathrm{Uniform}(0,1)$6 the sequence of skill contexts, and $U_i\sim \mathrm{Uniform}(0,1)$7 the environment context. Task feasibility is defined as the expected return of the current skill policy and predicted by a value function $U_i\sim \mathrm{Uniform}(0,1)$8. Task diversity is tied to entropy and approximated nonparametrically through the $U_i\sim \mathrm{Uniform}(0,1)$9-nearest-neighbor distance. ATR scores candidate tasks by

$1/g$0

or in embedding space

$1/g$1

with $1/g$2, and samples tasks via an $1/g$3-greedy rule that uses the prior with probability $1/g$4. The task encoder is a relational network jointly trained with the skill policies, and procedural generation is based on graph-structured parameters converted into simulated scenes (Fang et al., 2022).

ATR is evaluated on four skills—$1/g$5—and on three sequential manipulation tasks. The paper reports that ATR achieves comparable or better success rates than baselines in single-step skill learning, with the largest gains for place-onto and pull-with, and that ATR-trained skills outperform baselines on all three sequential manipulation tasks in both simulation and the real world. The reported real-world success rates range roughly from $1/g$6 to $1/g$7, and simulation success rates are around $1/g$8 to $1/g$9 (Fang et al., 2022).

In robotics, then, task-oriented randomization shifts attention away from indiscriminate perturbation. The randomized object is instead chosen because it is informative for the downstream controller, or because it defines a feasible and diverse curriculum for skill acquisition.

5. Randomization within model and pretraining pipelines

Task-oriented randomization also appears inside learning systems themselves. “A Tailored Pre-Training Model for Task-Oriented Dialog Generation” introduces Start Position Randomization (SPR) in PRAL, a GPT-2-based dialog pretraining model. GPT-2 uses absolute positional embeddings indexed from $1/g$0 to $1/g$1, and the paper identifies two problems with always starting a dialogue at position $1/g$2: most dialogs are much shorter than $1/g$3 tokens, leaving many positional vectors unused, and fixed early positions induce spurious associations between textual content and absolute positions. Let $1/g$4 denote the total number of tokens in a dialogue. SPR chooses the dialogue’s start index uniformly from

$1/g$5

and places the entire sequence starting from that offset. The method is therefore a data/positioning strategy rather than an auxiliary loss. Its purpose is to decouple positional index from textual meaning and to spread training signal across the full position-embedding range (Gu et al., 2020).

SPR is one of three modifications added on top of ARDM, together with teacher GPT / knowledge distillation and history discount. The full training objective is

$1/g$6

and the discounted language-modeling loss is

$1/g$7

SPR does not alter this formula directly; it changes how token positions are assigned before the loss is computed. In the CamRest676 ablation, full PRAL reports BLEU-4 $1/g$8 and Success F1 $1/g$9, whereas removing SPR gives BLEU-4 $g \mid N$00 and Success F1 $g \mid N$01. The same table shows that removing Teacher GPT has a larger effect, with BLEU-4 $g \mid N$02 and Success F1 $g \mid N$03, indicating that the three techniques are complementary rather than redundant (Gu et al., 2020).

A broader variant of the same design logic appears in “Stabilizing black-box algorithms through task-oriented randomization.” The framework replaces a single dataset $g \mid N$04 with $g \mid N$05 randomized datasets $g \mid N$06, feeds each through a black-box algorithm $g \mid N$07, and aggregates the outputs, typically by averaging,

$g \mid N$08

If the data-generating mechanism is known, the perturbation scheme is chosen to match it; the paper explicitly lists Gaussian, Laplace, exponential, bootstrap, and subsampling variants. If the mechanism is unknown, the paper turns to diffusion-based randomization with forward process

$g \mid N$09

and reverse model

$g \mid N$10

The paper defines $g \mid N$11-stability by leave-one-out sensitivity and states that

$g \mid N$12

implies $g \mid N$13-stability, where $g \mid N$14 measures instability under the noisification mechanism. It also derives a Hoeffding-type concentration bound for the deviation between the original algorithm and the noisified ensemble output (Wang et al., 24 Jun 2026).

The black-box stabilization paper emphasizes a stability–exploration trade-off. In the known-mechanism case, exploration is controlled by the variance of the added noise; in the unknown-mechanism case, it is related to total variation under diffusion. In simulation with $g \mid N$15 Gaussian-noisified datasets, moderate noise improves stability for the neural network when $g \mid N$16, whereas excessive noise harms stability when $g \mid N$17. On MNIST, the paper reports disagreement below $g \mid N$18 for every digit class, corresponding to less than $g \mid N$19 discrepancy with $g \mid N$20 test images per digit, and extends the framework conceptually to top-$g \mid N$21 ranking via aggregation of top-$g \mid N$22 outputs across randomized passes (Wang et al., 24 Jun 2026).

These examples show that task-oriented randomization need not act on treatment labels or sampling times. It can intervene at the level of positional indices, synthetic training tasks, or noisy input replicas, provided the randomization is matched to the computational task.

6. Guarantees, trade-offs, and limitations

The literature attaches formal guarantees to several of these task-specific mechanisms. Weighted randomization tests are intended to retain finite-sample validity under adaptive collection through design-aware weighting (Nair et al., 2023). The threshold-based clinical-trial method proves equal-probability statements over positions, treatments, and phases and guarantees exactly equal treatment numbers (Santos-Magalhaes et al., 2015). Sequential rerandomization and VNSRR connect improved balance to variance reduction and, under stated conditions, unbiased estimation of treatment effects (Zhou et al., 2017, Lu et al., 2023). The black-box stabilization framework gives $g \mid N$23-stability and concentration bounds, while HeartSteps evaluates average treatment and uniformity with GEE-based population-average analyses rather than relying only on simulation (Wang et al., 24 Jun 2026, Meng et al., 3 Jan 2025).

At the same time, these methods make clear that task-oriented randomization is not synonymous with uniformity or exchangeability. Some methods deliberately depart from uniform sampling over feasible assignments, as in VNSRR’s local search on the space of treatment vectors (Lu et al., 2023). Others replace exchangeability by weighted pathwise likelihoods (Nair et al., 2023). Others introduce nonuniform probabilities on purpose, as in SeqRTS, where the randomization probability must react online to remaining budget and estimated future opportunities (Meng et al., 3 Jan 2025). In robotics, the goal is often to avoid uniformly probing all parameters or uniformly sampling all tasks, because only a subset is task-relevant or feasible (Liang et al., 2020, Fang et al., 2022).

The limitations are correspondingly domain-specific. Weighted randomization testing requires a known or estimable adaptive assignment rule $g \mid N$24 and computationally tractable resampling algorithms (Nair et al., 2023). The strong-randomized-list procedure relies on concealment of the generated list; if the group order is revealed, subsequent phase assignments become deterministic (Santos-Magalhaes et al., 2015). Sequential rerandomization theory is strongest under large groups, approximate normality, and homogeneous covariance across groups, and the paper notes possible weaknesses with small groups, heavy tails, binary covariates, or substantial heterogeneity (Zhou et al., 2017). VNSRR relies on symmetry and initialization conditions for its unbiasedness result and still requires threshold specification through $g \mid N$25 or $g \mid N$26 (Lu et al., 2023).

The more adaptive and learned variants have additional practical fragilities. HeartSteps reports that SeqRTS met its constraints on average rather than for every participant or block, that a timezone conversion bug in $g \mid N$27 forced probabilities to the minimum $g \mid N$28 in some blocks, and that $g \mid N$29 overestimated remaining risk times by about $g \mid N$30 on average (Meng et al., 3 Jan 2025). Active task-oriented exploration assumes a known parameterized model, accessible observations sufficient for system identification, and incurs expensive nested optimization over exploration, identification, and planning (Liang et al., 2020). ATR assumes a predefined skill library and a hand-defined task planner, together with scene-graph construction heuristics (Fang et al., 2022). In PRAL, SPR improves results but has a smaller ablation effect than teacher-based distillation (Gu et al., 2020). In black-box stabilization, excessive noise weakens both stability and accuracy, and the top-$g \mid N$31 ranking extension is described as theoretically preliminary (Wang et al., 24 Jun 2026).

A plausible implication is that task-oriented randomization should be viewed less as a single method than as a methodology for aligning stochastic design with structure. The cited works collectively indicate that the success of the approach depends on whether the chosen randomization mechanism faithfully captures the constraints, sensitivities, and objectives of the task it is meant to serve.