AI Variance Budget Allocation

• AI Variance Budget is a framework that strategically reallocates fixed resources—such as samples, labels, and compute—to minimize estimator uncertainty in high-variance settings.
• It employs variance-aware allocation, directing more resources to high-variance arms, hard instances, or expensive predictors to reduce error, regret, and confidence interval width.
• This cross-domain approach spans methodologies from best-arm identification in bandits to adaptive label budgeting in LLM evaluations, optimizing performance under explicit constraints.

to=arxiv_search.search ￣第四色ાjson {"query":"(Lalitha et al., 2023) OR Fixed-Budget Best-Arm Identification with Heterogeneous Reward Variances", "max_results": 5, "sort_by":"relevance"}【อ่านข้อความเต็มanalysis to=arxiv_search.search code ＿俺去也json {"query":"(He et al., 15 Mar 2025) OR Variance-Dependent Regret Lower Bounds for Contextual Bandits", "max_results": 5, "sort_by":"relevance"} to=arxiv_search.search 天天中彩票粤json {"query":"(Saha et al., 17 Feb 2026) OR LLM-as-Judge on a Budget", "max_results": 5, "sort_by":"relevance"} to=arxiv_search.search 弘鼎json {"query":"(Brawand et al., 8 May 2026) OR Active Multiple-Prediction-Powered Inference", "max_results": 5, "sort_by":"relevance"} “AI Variance Budget” is a cross-domain research notion for allocating a fixed resource budget—samples, labels, judge queries, token compute, or multi-turn interaction steps—so that uncertainty is reduced where it matters most. Across recent work, the common mechanism is variance-aware allocation: high-variance arms, hard instances, expensive-but-informative predictors, or long-horizon trajectories receive disproportionate budget so that estimator variance, confidence interval width, regret, or worst-case estimation error is minimized under an explicit constraint (Lalitha et al., 2023, Saha et al., 17 Feb 2026, Brawand et al., 8 May 2026). The term does not denote a single formalism; rather, it names a family of optimization problems in which variance, residual uncertainty, censoring weight, or output dispersion becomes the quantity against which budget is traded.

1. Core allocation principle

At its most classical, the variance-budget principle equalizes estimator uncertainty across heterogeneous units. In fixed-budget best-arm identification with known arm variances, SHVar allocates stage budget so as to equalize $\sigma_i^2 / N_{s,i}$, with

$$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$

which is the per-stage $G$-optimal design (Lalitha et al., 2023). In worst-case score estimation for LLM-as-a-judge, the known-variance minimax allocation is

$$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$

because equalizing $\sigma_i^2/n_i$ minimizes the maximum standard error across pairs (Saha et al., 17 Feb 2026). In cost-aware weak/strong evaluation, the same logic appears in the Neyman-with-cost solution

$$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$

which balances weak-rater variance $V_f$, residual variance $V_r$, and annotation costs $c_w,c_s$ (Angelopoulos et al., 9 Jun 2025).

Setting	Budgeted quantity	Variance-aware target
Fixed-budget BAI	Per-stage pulls	$\sigma_i^2/N_{s,i}$
LLM-as-Judge worst-case estimation	Queries per pair	$$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$0
Weak/strong evaluation	Weak and strong labels	$$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$1
AM-PPI	Label propensity	$$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$2

A central distinction is objective dependence. When the objective is to minimize a maximum error, allocation is proportional to $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$3; when the objective is to minimize a sum of variances, classical Neyman allocation uses

$$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$4

which is explicitly suboptimal for the worst-case objective in LLM-as-a-judge (Saha et al., 17 Feb 2026). This suggests that an AI variance budget is not a universal formula but an objective-specific allocation rule.

2. Bandits, best-arm identification, and variance-dependent lower bounds

In fixed-budget best-arm identification, the budget is spent across arms before successive elimination. SHVar assumes known reward variances and pulls the arm maximizing $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$5; SHAdaVar replaces unknown variances by high-probability overestimates and, after a warm-up of $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$6 pulls per arm per stage, pulls the arm maximizing $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$7 (Lalitha et al., 2023). The resulting guarantees are exponential in budget. For SHVar,

$$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$8

and an alternative bound uses $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$9 instead of $G$0 (Lalitha et al., 2023). For SHAdaVar, under Gaussian noise and a global union bound,

$G$1

with $G$2 as $G$3 (Lalitha et al., 2023).

A related but distinct problem is variance-optimal arm selection, where the target is the arm with the largest variance. The paper introducing \texttt{UCB-VV} and \texttt{SHVV} defines variance regret

$G$4

and proves for bounded rewards that

$G$5

while \texttt{SHVV} achieves error probability of order $G$6, matching a lower bound up to constants (Khurshid et al., 17 May 2025). Here the variance budget is spent to discriminate arms by their variances rather than by their means.

In linear contextual bandits with heteroscedastic noise, the variance budget becomes the total variance

$G$7

The prefixed-sequence lower bound gives

$G$8

and the weak-adversary adaptive-sequence lower bound gives, with probability at least $G$9,

$$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$0

matching SAVE up to logarithmic factors (He et al., 15 Mar 2025). Yet the same paper proves that if the adversary chooses $$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$1 after seeing the decision set, there exists an algorithm with $$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$2 while $$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$3 (He et al., 15 Mar 2025). This directly refutes the misconception that a large total variance budget always implies a hard learning problem.

3. Budgeted evaluation with LLM judges, weak raters, and strong raters

In LLM-as-a-judge, each prompt–response pair $$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$4 has a stochastic score

$$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$5

and the objective is to minimize

$$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$6

under a fixed query budget $$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$7 (Saha et al., 17 Feb 2026). With known variances, equalizing $$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$8 yields high-probability worst-case error

$$n_i^\star = B \,\frac{\sigma_i^2}{\sum_{j=1}^K \sigma_j^2},$$9

with probability at least $\sigma_i^2/n_i$0 (Saha et al., 17 Feb 2026). In the unknown-variance algorithm, a pilot of

$\sigma_i^2/n_i$1

queries per pair supports the variance-UCB

$\sigma_i^2/n_i$2

and the adaptive rule $\sigma_i^2/n_i$3 achieves the same $\sigma_i^2/n_i$4 rate up to logarithmic factors (Saha et al., 17 Feb 2026). On HelpSteer2, the reported “Half-budget savings” means that the worst-case estimation error achieved by the adaptive method at $\sigma_i^2/n_i$5 is comparable to uniform allocation at $\sigma_i^2/n_i$6 (Saha et al., 17 Feb 2026).

When there are multiple judges with different costs, the problem becomes budgeted heteroskedastic multi-judge estimation. The inverse-variance weighted estimator is

$\sigma_i^2/n_i$7

and the oracle allocation is sparse: for each instance $\sigma_i^2/n_i$8, all budget goes to

$\sigma_i^2/n_i$9

with

$$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$0

for $$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$1 error (Lee et al., 22 May 2026). Est-IVWE uses forced exploration, optimistic variances $$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$2, and empirical oracle allocation; its leading error term matches the oracle rate up to lower-order terms, and a local Assouad-type minimax lower bound shows the sharp dependence on $$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$3 (Lee et al., 22 May 2026).

A related evaluation problem uses a cheap weak rater and an expensive strong rater. The batch PPI mean estimator

$$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$4

is unbiased for $$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$5, with approximate variance

$$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$6

and cost-optimal ratio

$$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$7

(Angelopoulos et al., 9 Jun 2025). Under heteroskedastic residuals $$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$8, the optimal active policy has clipped square-root form

$$\frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}}, \qquad \mathrm{Var}_{\min}(\hat{\mu}; B) = \left(\frac{\sqrt{c_w V_f} + \sqrt{c_s V_r}}{B}\right)^2,$$9

so strong labels are concentrated on hard examples (Angelopoulos et al., 9 Jun 2025). Reported real-data gains include reaching RMSE $V_f$0 at roughly $V_f$1 of the budget of the human-only baseline in the overall Chatbot Arena slice, and $V_f$2 in an “easy+hard” slice (Angelopoulos et al., 9 Jun 2025).

4. Active statistical inference, preferences, and randomized experiments

In post-deployment monitoring, the budget must often be spent jointly on predictor queries and clinician labels. AM-PPI formalizes this with the estimator

$V_f$3

whose asymptotic variance is

$V_f$4

The optimization problem is

$V_f$5

and the KKT conditions yield the proportional-to-residual-uncertainty rule

$V_f$6

clipped at $V_f$7, together with weighted least-squares reweighting and per-instance routing (Brawand et al., 8 May 2026). The paper proves biconvexity with closed-form partial minimizers, strong duality, asymptotic normality, and minimum-variance unbiasedness within the linear-prediction AIPW class, and reports $V_f$8 to $V_f$9 percent narrower confidence intervals than single-predictor ASI in the routing-relevant regime (Brawand et al., 8 May 2026).

Budget-constrained acquisition can also trade ground-truth labels against pairwise preferences. PCAL casts this as monotone missingness with observation patterns $V_r$0 and budget constraint

$V_r$1

Its asymptotic variance objective is minimized over acquisition policies $V_r$2, and under MCAR the paper states that if

$V_r$3

then the variance-minimizing allocation uses a nonzero $V_r$4, meaning preferences should be purchased (Dong et al., 19 Jan 2026). The estimator is asymptotically normal, semiparametrically efficient for the chosen policy, and enjoys the robustness guarantee

$V_r$5

in trace-variance, asymptotically, up to small tolerances (Dong et al., 19 Jan 2026).

In randomized experiments, AI-generated predictions enter as prognostic covariates rather than replacement outcomes. The recommended estimator is Lin-style regression adjustment,

$V_r$6

with AI scores included in $V_r$7 (Arbour et al., 7 Jun 2026). Under complete randomization, the difference-in-means variance is

$V_r$8

and regression adjustment reduces the within-arm component by the prognostic correlation, with arm-specific design effect

$V_r$9

The paper emphasizes a “do no harm” property: if AI scores are uninformative, $c_w,c_s$0 reverts to $c_w,c_s$1, whereas an uncalibrated model-assisted estimator has $c_w,c_s$2 and inflates variance whenever $c_w,c_s$3 (Arbour et al., 7 Jun 2026).

5. Compute-time allocation, dynamic censoring, and generative dispersion

Variance budgeting also appears in reinforcement learning for anytime reasoning. AnytimeReasoner defines

$c_w,c_s$4

where budgets $c_w,c_s$5 are sampled from a prior and dense verifiable rewards are produced by summarizing truncated reasoning traces (Qi et al., 19 May 2025). For the thinking policy, BRPO replaces the GRPO group-average baseline with an interpolation

$c_w,c_s$6

where $c_w,c_s$7 uses earlier-budget verified rewards on the same prefix and $c_w,c_s$8 is a group-average baseline (Qi et al., 19 May 2025). The paper reports lower normalized variance than GRPO, especially for long thinking, and consistent improvements across all thinking budgets. At max budget $c_w,c_s$9, “AnytimeReasoner-base” reaches average $\sigma_i^2/N_{s,i}$0 versus $\sigma_i^2/N_{s,i}$1 for GRPO; the anytime average is best for “AR-uniform” at $\sigma_i^2/N_{s,i}$2 versus $\sigma_i^2/N_{s,i}$3 for GRPO (Qi et al., 19 May 2025).

For multi-turn evaluation, DAPRO makes the budget explicit through

$\sigma_i^2/N_{s,i}$4

where $\sigma_i^2/N_{s,i}$5 is a censored time-to-event (Feldman et al., 7 May 2026). Phase I solves

$\sigma_i^2/N_{s,i}$6

thereby minimizing the mean inverse censoring probability under a budget constraint (Feldman et al., 7 May 2026). The resulting coverage gap depends on the square root of the mean censoring weight,

$\sigma_i^2/N_{s,i}$7

rather than a worst-case weight, and the same dynamic allocation supports unbiased, low-variance IPW estimates of metrics such as jailbreak rate and restricted mean time-to-event (Feldman et al., 7 May 2026).

A different use of “variance” concerns deliberate output dispersion in brainstorming. “Prompting Diverse Ideas” defines diversity through mean pairwise cosine similarity, number of unique ideas, and exhaustion curves (Meincke et al., 2024). In that setting, the “variance budget” is spent through prompt design and sampling protocols rather than through statistical estimation. The reported average cosine is $\sigma_i^2/N_{s,i}$8 for the base prompt, $\sigma_i^2/N_{s,i}$9 for Chain-of-Thought, and $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$00 for the human group benchmark; the CoT opportunity space rises to approximately $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$01 unique ideas versus approximately $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$02 for the base prompt, with duplicate shares of approximately $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$03 versus approximately $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$04 (Meincke et al., 2024). This suggests a broader interpretation in which variance is a resource for exploration of idea space, but the paper also reports that the CoT advantage becomes negligible after roughly $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$05–$$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$06 ideas as the accessible pool is depleted (Meincke et al., 2024).

6. Limits, adversaries, environment effects, and open questions

A recurring limitation is distributional dependence. SHAdaVar’s theoretical guarantees use Gaussian noise and chi-square concentration, though the paper notes empirical robustness on MovieLens and points to empirical Bernstein extensions for non-Gaussian noise (Lalitha et al., 2023). Multi-judge estimation assumes unbiased judges with bounded scores and independent samples; judge correlations or systematic biases fall outside the model and would require covariance-aware or debiased estimators (Lee et al., 22 May 2026). Cost-optimal active evaluation likewise assumes MAR conditional on $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$07 and can lose efficiency under policy misspecification, although the paper provides a variance inflation bound under inverse-propensity error (Angelopoulos et al., 9 Jun 2025).

Another limitation is that a variance budget can be nullified by the wrong adversary model. In linear contextual bandits, variance-dependent lower bounds hold for prefixed sequences and for adaptive sequences when the adversary must choose $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$08 before seeing $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$09, but they fail under the strong-adversary timing model (He et al., 15 Mar 2025). In randomized experiments, the “do no harm” guarantee belongs to regression adjustment with centered covariates and interactions; the same paper states that model-assisted prediction substitution lacks this property and can increase variance (Arbour et al., 7 Jun 2026). A plausible implication is that variance budgets are meaningful only relative to a specified estimator class and information structure.

System-level variance budgets arise even when the stochastic object is not a statistical estimator but the runtime environment. A study of $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$10 open-source AI-enabled systems ran $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$11 repetitions in each of eight environment configurations and measured model performance, processing time, and expense (Rahman et al., 2024). Between Linux and MacOS, statistically significant instability was observed in $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$12, $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$13, and $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$14 of the studied projects for model performance, processing time, and expense, respectively; between Linux and Windows, the corresponding values were $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$15, $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$16, and $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$17 (Rahman et al., 2024). The paper’s practical recommendation is Linux + AMD64 + Python 3.7 as the default stable choice (Rahman et al., 2024). This is a distinct sense of variance budgeting: environment selection reduces exogenous variance before any downstream learning or evaluation budget is spent.

At the lower-bound level, continuous simulation optimization shows a “variance dichotomy.” The minimax regret lower bound is the maximum of a variance-dependent term and a variance-independent term: $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$18 for $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$19, and

$$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$20

for $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$21 (Du et al., 15 Apr 2026). The threshold $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$22 for $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$23 separates a deterministic-like regime, where variance reduction has little payoff, from a stochastic regime, where reducing $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$24 matters directly (Du et al., 15 Apr 2026). This lower-bound perspective sharpens a general lesson visible across the literature: spending budget on variance reduction is rational only after the variance-independent bottleneck has been passed.

Across these literatures, “AI Variance Budget” therefore denotes a technical design doctrine rather than a single method. It includes $$n_{s,i} = \left( \frac{\sigma_i^2}{\sum_{j\in A_s} \sigma_j^2} \right) n_s,$$25-optimal arm pulls, minimax query allocation, clipped square-root label propensities, weighted inverse-propensity estimators, budget-relative control variates, and even prompt engineering for deliberate dispersion. What unifies these uses is the same structural question: under a finite budget, which units, stages, examples, predictors, judges, or prefixes should absorb additional variance-reduction effort so that the resulting estimator, confidence interval, regret, or exploration process is as effective as possible?