Budgeted Deferral in Decision Systems

• Budgeted deferral is a resource allocation principle where systems judiciously defer decisions to stronger or more reliable experts when necessary.
• It is applied across diverse settings such as cascaded machine translation, online human deferral, and workload shifting, each utilizing distinct budget representations.
• Methodologies range from threshold-based rules to cost-sensitive classifiers, balancing performance benefits against resource limitations and intervention costs.

Budgeted deferral denotes a family of decision-making schemes in which a system does not invoke the strongest, most expensive, or most reliable option uniformly, but instead allocates scarce intervention capacity selectively. In recent work, the governing resource can be the deferred fraction $\eta$ in a model cascade, a hard cumulative cap on human or update actions, explicit expert-specific costs embedded in the loss, the number of consulted entities, deadline slack expressed as a maximum postponement window, or the number of expert queries spent during training (Farinhas et al., 18 Feb 2025, Reid et al., 2024, Wang, 23 Jun 2026, DeSalvo et al., 30 Oct 2025, Adnan et al., 2011).

1. Conceptual scope

In the most familiar learning-to-defer formulation, a predictor decides whether to act autonomously or defer to an expert. The recent literature broadens this basic picture in several directions. In online decision deferral, the model arm is free and human deferral incurs heterogeneous cost $c_t$, so the central problem is whether the expected gain from deferring justifies spending scarce expert budget on that round (Reid et al., 2024). In multi-expert routing, the action is no longer binary: the learner may defer to one of several fixed experts with different costs and error profiles, or consult multiple entities under Top-$k$ selection (Cortes et al., 30 Apr 2026, Montreuil et al., 15 May 2025). In regression, the same structure reappears with a continuous label space, where the system either predicts with $h(x)$ or routes to one of several experts and incurs expert-specific cost $c_j(x,y)$ (Mao et al., 2024).

Other papers defer something other than the prediction itself. ADOWIP studies delayed online forecasting in which the scarce intervention is the update rather than the prediction: at each revealed feedback event $i$, the learner chooses $g_i\in\{0,1\}$, and the hard budget is $\sum_i g_i c_i \le B$ (Wang, 23 Jun 2026). The data-center literature treats workload execution as the deferred object: a workload unit released at time $t$ may be shifted to $t+d$, but only up to a deadline $c_t$0, so the “budget” is a hard postponement window rather than a review quota (Adnan et al., 2011, Adnan et al., 2012). DoorDash’s pacing work similarly treats ad-auction participation as a temporally deferred spend decision over a fixed day, using throttling to preserve budget for later requests (Garg et al., 9 Sep 2025).

This suggests that “budgeted deferral” is better understood as a resource-allocation principle than as a single formalism. What remains invariant is the presence of a fallback action that is stronger, safer, or more appropriate on some subset of cases, together with a resource notion that prevents universal fallback.

2. Budget representations

The literature uses several non-equivalent budget representations.

Setting	What is deferred	Budget representation
Cascaded MT	selected examples	deferred fraction $c_t$1
Online human deferral	human escalations	$c_t$2
Delayed online adaptation	accepted updates	$c_t$3
Top-$c_t$4 L2D	consulted entities	fixed $c_t$5 or adaptive $c_t$6
Multi-expert medical routing	average deferral mass	$c_t$7
Workload shifting	execution over time	deadline window $c_t$8

In cascaded machine translation, the budget is operationalized as a target deferral rate: every sentence is first translated by the small model, hypotheses are scored by a QE metric, and the lowest-scoring $c_t$9 fraction is re-run through the large model. The paper explicitly adopts a budget-constrained computation approach rather than a fixed absolute threshold, with the full test set for each language pair treated as a single batch in the reported WMT24 experiments (Farinhas et al., 18 Feb 2025). In sequential LLM agents, ReDAct calibrates an uncertainty threshold $k$0 so that the average number of large-model calls per episode is close to a target $k$1, which functions as a proxy budget on expensive decisions rather than a direct token-cost constraint (Piatrashyn et al., 8 Apr 2026).

Several papers instead impose explicit cumulative constraints. In online decision deferral, the learner must satisfy $k$2, where the human action is costly and the model action has zero cost (Reid et al., 2024). ADOWIP uses the same hard-budget form, $k$3, but over accepted updates after delayed feedback revelation (Wang, 23 Jun 2026). MPD$k$4-Router uses a dataset-level soft average budget, $k$5, enforced by a one-sided augmented Lagrangian rather than a hard realized cap (Zhan, 8 May 2026).

A different line of work embeds the budget into the objective rather than the feasible set. MILD optimizes expected deferral loss with expert-dependent costs such as $k$6; the $k$7 terms act as resource prices and yield what the paper explicitly characterizes as a soft-budgeted, Lagrangian-style routing formulation (Cortes et al., 30 Apr 2026). The same interpretation appears in regression with multi-expert deferral, where $k$8 or other bounded costs can encode the price of expert use (Mao et al., 2024).

Deadline-based work encodes budget as maximum tolerated delay. In capacity provisioning and geographical load balancing, a release at slot $k$9 must execute by $h(x)$0, or by a class-specific deadline $h(x)$1 in the nonuniform case, so the budget is a hard temporal slack variable rather than a deferral fraction (Adnan et al., 2011, Adnan et al., 2012). Training-time budgeted deferral introduces yet another resource: the number of expert-cost queries used while learning the defer policy, replacing the full-information $h(x)$2 supervision pattern by selective querying with importance weighting (DeSalvo et al., 30 Oct 2025).

3. Decision rules and routing mechanisms

The simplest mechanisms are threshold or ranking rules. In cascaded translation, the decision rule is ranking-based: compute a QE score $h(x)$3 for each small-model translation, sort by score, and defer the bottom $h(x)$4 fraction. The paper emphasizes that this avoids universal score calibration and instead performs budget calibration by batch quantile (Farinhas et al., 18 Feb 2025). ReDAct uses a pointwise uncertainty threshold: at each time step, the small model proposes an action, computes an uncertainty score $h(x)$5 from action-token probabilities, and defers when $h(x)$6. The threshold is calibrated on held-out episodes to target a desired average number of expensive calls (Piatrashyn et al., 8 Apr 2026). Density-ratio post-hoc L2D learns a scalar scorer once and then defers by thresholding that scorer, which allows the deferral rate to be adjusted without retraining (Soen et al., 19 May 2026).

A second family learns the router directly as a cost-sensitive classifier. MILD recasts two-stage multi-expert deferral as cost-sensitive classification over the input-expert domain, with routing score $h(x)$7, margin $h(x)$8, and a cost-sensitive surrogate in which minority experts receive their own margin parameters $h(x)$9 (Cortes et al., 30 Apr 2026). One-stage Top-$c_j(x,y)$0 L2D generalizes this score-based picture to selecting the $c_j(x,y)$1 highest-scoring entities from a joint label-and-expert set $c_j(x,y)$2, using a surrogate that is independent of $c_j(x,y)$3, so a single trained model can be deployed under different consultation budgets (Montreuil et al., 15 May 2025). The dissertation on abstention and multi-expert deferral systematizes both score-based and predictor-router constructions across single-stage and two-stage settings, including the case where abstention is treated as deferral to a single expert (Mao, 28 Dec 2025).

Other mechanisms are structured around operational constraints rather than pure scoring. MPD$c_j(x,y)$4-Router factorizes the policy into a defer head $c_j(x,y)$5 and a conditional expert-allocation distribution $c_j(x,y)$6, then uses mask-aware Gumbel-sigmoid support selection and masked softmax allocation so that unavailable experts receive zero probability exactly (Zhan, 8 May 2026). DeferredSeg makes the routing decision at pixel level: for each pixel, a softmax over $c_j(x,y)$7 branches decides whether the base segmentor keeps responsibility or a specific expert branch receives that pixel, with an aggregated deferral predictor promoting smoother regions (Tian et al., 14 Apr 2026). ADOWIP uses an observed decision-loss priority gate, updating only when

$c_j(x,y)$8

exceeds both a cost-scaled threshold and an empirical quantile of past revealed scores, provided budget remains (Wang, 23 Jun 2026).

In sequential and online settings, dual or recourse variables often replace static thresholds. The contextual-bandit deferral algorithm chooses between model and human by maximizing optimistic reward minus a dual-weighted optimistic cost,

$c_j(x,y)$9

where $i$0 tracks budget pressure (Reid et al., 2024). In Smart Fast Finish, temporal deferral is controlled by a campaign-specific late-day start time and a transition window that gradually reduces throttle rate to zero, rather than by a per-request classifier (Garg et al., 9 Sep 2025). In deadline-constrained workload shifting, the deferral rule is encoded by variables $i$1: choosing $i$2 runs now, while $i$3 consumes deadline slack (Adnan et al., 2011, Adnan et al., 2012).

4. Optimization principles and guarantees

A central theoretical divide runs between hard-feasibility methods and penalty-based methods. Hard-feasibility results appear in contextual bandits with knapsacks, delayed-update control, and deadline-constrained scheduling. Online decision deferral proves a high-probability regret bound against an optimal static policy under $i$4, assuming $i$5 and generalized linear reward and cost models (Reid et al., 2024). ADOWIP proves exact hard-budget feasibility—if the budget object refuses spends above $i$6, then $i$7—along with projected-OGD regret for the convex accepted-update subproblem, scheduler stability near the quantile threshold, and conditional finite-sample threshold-selection statements (Wang, 23 Jun 2026). In capacity provisioning, VFW$i$8 and GCP guarantee feasibility under hard deadlines and admit constant-factor worst-case bounds relative to the offline LP (Adnan et al., 2011).

Penalty-based methods instead optimize a surrogate risk in which expert use carries cost. MILD derives new margin-based surrogates and $i$9-consistency bounds for imbalanced multi-expert deferral, showing that low surrogate regret implies low regret in the original cost-sensitive deferral objective (Cortes et al., 30 Apr 2026). Regression with multi-expert deferral provides analogous single-stage and two-stage surrogate constructions with $g_i\in\{0,1\}$0-consistency bounds for bounded regression losses and bounded expert costs (Mao et al., 2024). The broader dissertation extends these ideas across multi-class abstention, multi-expert classification deferral, and regression, with non-asymptotic, hypothesis-set-specific consistency guarantees and realizable consistency results in several two-stage constant-cost settings (Mao, 28 Dec 2025). Top-$g_i\in\{0,1\}$1 one-stage L2D adds Bayes-consistency and $g_i\in\{0,1\}$2-consistency for a convex surrogate independent of $g_i\in\{0,1\}$3 (Montreuil et al., 15 May 2025).

Budgeted deferral can also target training rather than deployment. The training-time framework of “Budgeted Multiple-Expert Deferral” replaces full expert supervision by selective expert-cost querying, uses an unbiased importance-weighted empirical risk, and gives generalization bounds plus label-complexity analyses. In the realizable case, the expected number of expert-cost queries scales as $g_i\in\{0,1\}$4 in the main analysis and can be sharpened to logarithmic in $g_i\in\{0,1\}$5 via Freedman-style arguments (DeSalvo et al., 30 Oct 2025).

Several papers provide explicit resource accounting. Cascaded MT formalizes compute with a transformer FLOPs approximation and reports that, in the Tower-v2 7B $g_i\in\{0,1\}$6 70B setup, compute parity occurs at $g_i\in\{0,1\}$7 deferral with wmt22-cometkiwi-da and $g_i\in\{0,1\}$8 with wmt23-cometkiwi-da-xxl, so the empirically useful operating points remain below always-large cost (Farinhas et al., 18 Feb 2025). In post-hoc density-ratio L2D, the practical guarantee is different: train a single scorer and choose the threshold afterward, turning budget selection into a deployment-time quantile problem rather than a training-time constraint (Soen et al., 19 May 2026).

5. Empirical regimes and application domains

Machine translation provides one of the clearest empirical demonstrations of budgeted deferral. A two-stage QE-based cascade can match Tower-v2 70B while invoking it on only $g_i\in\{0,1\}$9 to $\sum_i g_i c_i \le B$0 of examples in automatic evaluation, and is not significantly different from the 70B model while deferring only about $\sum_i g_i c_i \le B$1 to $\sum_i g_i c_i \le B$2 of examples in human evaluation on en-es and en-ja (Farinhas et al., 18 Feb 2025). Sequential LLM agents show a similar pattern: ReDAct reports that deferring only about $\sum_i g_i c_i \le B$3 of decisions to the large model can match the quality of using it exclusively in some settings, while substantially reducing inference cost (Piatrashyn et al., 8 Apr 2026). In cost-aware LLM routing on MMLU, MILD avoids collapse to either the strongest or cheapest expert and achieves a markedly better deferral loss than a prior two-stage baseline in the error-plus-cost regime (Cortes et al., 30 Apr 2026).

Budgeted deferral is equally central in online and control settings. ADOWIP reports that, on held-out ETT capacity-planning splits, the selected gate yields 20/0 wins against exact-update always, fixed-period, and drift-triggered baselines under matched compute, and that the same protocol improves the UCI Bike capacity proxy with 20/0 held-out wins (Wang, 23 Jun 2026). In online human-AI deferral, the contextual-bandit approach approaches the empirical optimum on synthetic data and outperforms simple baselines on real knapsack and ImageNet16H tasks, especially when human benefit cannot be inferred from model confidence alone (Reid et al., 2024).

Medical and safety-critical routing has produced more structured multi-expert forms. DeferredSeg reports consistent improvements over baselines on PROMISE12, LiTS, and Chaksu, with deferral concentrating around boundaries and low-contrast regions, and with load-balancing penalties becoming important once several experts are introduced (Tian et al., 14 Apr 2026). MPD$\sum_i g_i c_i \le B$4-Router shows monotone, conservative deferral-rate control under augmented-Lagrangian targets: increasing $\sum_i g_i c_i \le B$5 increases realized deferral, but realized soft and hard rates remain below target, while moderate deferral budgets recover most of the unconstrained performance gains (Zhan, 8 May 2026).

Industrial pacing and systems scheduling show a different empirical profile. Smart Fast Finish increases campaign live hours by $\sum_i g_i c_i \le B$6 minutes overall and reduces overdelivery from $\sum_i g_i c_i \le B$7 to $\sum_i g_i c_i \le B$8, with larger gains on the largest campaigns (Garg et al., 9 Sep 2025). Dynamic deferral of workload yields $\sum_i g_i c_i \le B$9–$t$0 cost savings in data-center capacity provisioning and $t$1–$t$2 cost savings in geographical load balancing, with the benefit increasing as deadline slack increases (Adnan et al., 2011, Adnan et al., 2012).

6. Limitations, misconceptions, and open directions

A common misconception is that budgeted deferral is synonymous with uncertainty thresholding. The literature is much broader. Some methods do threshold uncertainty, such as ReDAct; others rank by QE scores, learn cost-sensitive routers, use density-ratio scorers, optimize primal-dual policies, or solve deadline-constrained linear programs (Piatrashyn et al., 8 Apr 2026, Farinhas et al., 18 Feb 2025, Cortes et al., 30 Apr 2026, Soen et al., 19 May 2026, Reid et al., 2024, Adnan et al., 2011).

A second misconception is that all methods provide exact budget guarantees. Many do not. Cost-sensitive and post-hoc thresholded methods provide soft or validation-calibrated control rather than guaranteed realized caps; this is explicit in soft-cost routing, score-threshold deferral, and augmented-Lagrangian average-budget control (Cortes et al., 30 Apr 2026, Mao et al., 2024, Soen et al., 19 May 2026, Zhan, 8 May 2026). By contrast, hard-feasibility statements appear only in settings where the budget is enforced by construction, such as $t$3, $t$4, or hard deadlines $t$5 (Reid et al., 2024, Wang, 23 Jun 2026, Adnan et al., 2012).

Calibration and shift remain persistent difficulties. Cascaded MT avoids universal thresholds precisely because absolute QE calibration is difficult, but this ties the method to batch-level score distributions and leaves dynamic thresholding for future work (Farinhas et al., 18 Feb 2025). ReDAct calibrates on small-model-only trajectories, yet deployment changes future states once deferral is active, producing a discrepancy between target and realized expensive-call counts (Piatrashyn et al., 8 Apr 2026). Post-hoc density-ratio L2D likewise relies on validation-based threshold selection, so the achieved budget under distribution shift depends on how representative validation is (Soen et al., 19 May 2026).

Several frameworks depend critically on the quality of the fallback or auxiliary signal. QE-based translation deferral works because MT has strong reference-free QE models and assumes the small model is reasonably competitive with the large one (Farinhas et al., 18 Feb 2025). ADOWIP assumes that revealed downstream decision loss is observable and informative about future update value (Wang, 23 Jun 2026). DeferredSeg requires expert masks during training and introduces only soft deferral-rate control, not a hard global referral budget (Tian et al., 14 Apr 2026). MPD$t$6-Router models availability exactly but expert capacities only indirectly, through priors and anti-collapse regularization rather than enforceable quotas (Zhan, 8 May 2026). Training-time budgeted deferral, finally, addresses expert-query cost during learning, which is a different budget axis from inference-time escalation (DeSalvo et al., 30 Oct 2025).

Open directions recur across the literature. These include dynamic thresholding outside batch settings, multistage cascades beyond two-stage designs, direct optimization of token-cost rather than call-count budgets, hard expert-capacity constraints in multi-expert routing, and stronger robustness analyses under domain shift (Farinhas et al., 18 Feb 2025, Piatrashyn et al., 8 Apr 2026, Zhan, 8 May 2026). A plausible synthesis is that future work will continue to move from scalar defer-or-predict decisions toward richer constrained allocation problems in which the system must decide not only whether to defer, but also when, to whom, for how long, and under which resource accounting regime.