--- title: AI Variance Budget Allocation url: https://www.emergentmind.com/topics/ai-variance-budget type: topic --- # AI Variance Budget Allocation to=arxiv_search.search ￣第四色ાjson {"query":"2306.07549 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":"2503.12020 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":"2602.15481 OR LLM-as-Judge on a Budget", "max_results": 5, "sort_by":"relevance"} to=arxiv_search.search 弘鼎json {"query":"2605.08429 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 [2306.07549] [2602.15481] [2605.08429]. 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 [2306.07549]. 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 [2602.15481]. 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 $c_w,c_s$ [2506.07949]. | 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 | $n_i^\star = B \sigma_i^2 / \sum_j \sigma_j^2$ | | Weak/strong evaluation | Weak and strong labels | $m^*/n^* = \sqrt{V_r c_w /(V_f c_s)}$ | | AM-PPI | Label propensity | $\pi_I^*(x) = \sqrt{r_I(x)/(n\,\mu\,c_{\text{label}})}$ | A central distinction is objective dependence. When the objective is to minimize a maximum error, allocation is proportional to $\sigma_i^2$; when the objective is to minimize a sum of variances, classical Neyman allocation uses \[ n_i^{\text{Neyman}} = B\, \frac{\sigma_i}{\sum_{j=1}^K \sigma_j}, \] which is explicitly suboptimal for the worst-case objective in LLM-as-a-judge [2602.15481]. 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 $\sigma_i^2/N_{s,t,i}$; SHAdaVar replaces unknown variances by high-probability overestimates and, after a warm-up of $4\log(1/\delta)+1$ pulls per arm per stage, pulls the arm maximizing $U_{s,t,i}/N_{s,t,i}$ [2306.07549]. The resulting guarantees are exponential in budget. For SHVar, \[ P\{\hat I \neq 1\} \le 2 \log_2 K \exp\!\left[ - \frac{n \, \Delta_{\min}^2}{4 \log_2 K \sum_{j \in A} \sigma_j^2} \right], \] and an alternative bound uses $\sigma_{\max}^2$ instead of $\sum_j \sigma_j^2$ [2306.07549]. For SHAdaVar, under Gaussian noise and a global union bound, \[ P\{\hat I \neq 1\} \le 2 \log_2 K \exp\!\left[ - \alpha \, \frac{(n - K \log K) \, \Delta_{\min}^2}{4 \, \sigma_{\max}^2 \, K \, \log_2 K} \right], \] with $\alpha \to 1$ as $n \to \infty$ [2306.07549]. 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 \[ E[R(n)] = \sum_{i=1}^K E[s_i(n)] \delta_i, \] and proves for bounded rewards that \[ R_{VV}(n) \le 8 \sum_{i: \sigma_i^2 < \sigma_*^2} \frac{\log n}{\delta_i} + \left(1 + \frac{\pi^2}{3}\right) \sum_{i: \sigma_i^2 < \sigma_*^2} \delta_i, \] while \texttt{SHVV} achieves error probability of order $\exp(-n/(\log K \cdot H))$, matching a lower bound up to constants [2505.11985]. 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 \[ \Lambda := \sum_{k=1}^K \sigma_k^2. \] The prefixed-sequence lower bound gives \[ E[\mathrm{Regret}(K)] \ge \Omega\!\left( d \sqrt{\sum_{k=1}^K \sigma_k^2} / \log K \right), \] and the weak-adversary adaptive-sequence lower bound gives, with probability at least $1-1/K$, \[ \mathrm{Regret}(K) \ge \Omega\!\left( d\sqrt{ \sum_{k=1}^K \sigma_k^2} /\log^6(dK) \right), \] matching SAVE up to logarithmic factors [2503.12020]. Yet the same paper proves that if the adversary chooses $\sigma_k^2$ after seeing the decision set, there exists an algorithm with $\mathrm{Regret}(K)\le d$ while $\sum_{k=1}^K \sigma_k^2 \ge K/2$ [2503.12020]. 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 $i$ has a stochastic score \[ X_{i,j} = s_i + \epsilon_{i,j}, \qquad \hat{s}_i = \frac{1}{n_i} \sum_{j=1}^{n_i} X_{i,j}, \] and the objective is to minimize \[ \|s - \hat{s}\|_\infty = \max_{i \in [K]} |s_i - \hat{s}_i| \] under a fixed query budget $B$ [2602.15481]. With known variances, equalizing $\sigma_i^2/n_i$ yields high-probability worst-case error \[ \max_{i \in [K]} |s_i - \hat{s}_i| \le \sqrt{\frac{\sum_{i=1}^K \sigma_i^2}{B}\,\log\frac{2K}{\delta}} \] with probability at least $1-\delta$ [2602.15481]. In the unknown-variance algorithm, a pilot of \[ t_0 = 16\log\left(\frac{4 K B}{\delta}\right) \] queries per pair supports the variance-UCB \[ U_i(t) := \frac{\hat{\sigma}_i^2(t)}{1 - \sqrt{\frac{4\log\left(4 K B / \delta\right)}{n_i(t)}}}, \] and the adaptive rule $i_t \in \arg\max_i U_i(t-1)/n_i(t-1)$ achieves the same $\tilde O\!\left(\sqrt{\sum_i \sigma_i^2/B}\right)$ rate up to logarithmic factors [2602.15481]. On HelpSteer2, the reported “Half-budget savings” means that the worst-case estimation error achieved by the adaptive method at $B=50k$ is comparable to uniform allocation at $B=100k$ [2602.15481]. When there are multiple judges with different costs, the problem becomes budgeted heteroskedastic multi-judge estimation. The inverse-variance weighted estimator is \[ \hat{s}_k = \left( \sum_{j=1}^J \frac{N_{k,j}}{\sigma_{k,j}^2} \right)^{-1} \sum_{j=1}^J \frac{N_{k,j} \hat{s}_{k,j}}{\sigma_{k,j}^2}, \] and the oracle allocation is sparse: for each instance $k$, all budget goes to \[ j^*(k) = \arg\min_j c_j \sigma_{k,j}^2, \] with \[ \omega_{k,j}^* = \frac{\mathbf{1}[j = j^*(k)] \big( c_{j^*(k)} \sigma_{k,j^*(k)}^2 \big)^{\frac{p}{p+2}}}{\sum_{k' \in [K]} \big( c_{j^*(k')} \sigma_{k',j^*(k')}^2 \big)^{\frac{p}{p+2}}} \] for $\ell_p$ error [2605.23362]. Est-IVWE uses forced exploration, optimistic variances $\overline{\sigma}_{k,j}=\hat{\sigma}_{k,j}+\tau$, 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 $\mathcal{A}_p^*(\bm\sigma,\mathbf c)$ [2605.23362]. A related evaluation problem uses a cheap weak rater and an expensive strong rater. The batch PPI mean estimator \[ \hat{\mu}_{\mathrm{PPI}} = \frac{1}{n}\sum_{i=1}^n f(X_i) \;+\; \frac{1}{m}\sum_{i\in L}\big(Y_i^* - f(X_i)\big) \] is unbiased for $\mu=\mathbb E[Y^*]$, with approximate variance \[ \mathrm{Var}(\hat{\mu}) \approx \frac{V_f}{n} + \frac{V_r}{m} \] and cost-optimal ratio \[ \frac{m^*}{n^*} = \sqrt{\frac{V_r\,c_w}{V_f\,c_s}} \] [2506.07949]. Under heteroskedastic residuals $u(x)=\mathbb E[(Y^*-f(X))^2\mid X=x]$, the optimal active policy has clipped square-root form \[ \pi_{\mathrm{active}}(x) = \min\!\big(\,\gamma^*(\tau)\,\sqrt{u(x)}\,,\,1\,\big), \] so strong labels are concentrated on hard examples [2506.07949]. Reported real-data gains include reaching RMSE $\approx 0.05$ at roughly $40\%$ of the budget of the human-only baseline in the overall Chatbot Arena slice, and $\approx 50\%$ in an “easy+hard” slice [2506.07949]. ## 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 \[ \hat{\theta} = \frac{1}{n} \sum_{i=1}^{n} \left[ \lambda_{I_i}^\top f_{I_i}(X_i) + \left( Y_i - \lambda_{I_i}^\top f_{I_i}(X_i) \right) \frac{\xi_i}{\pi_{I_i}(X_i)} \right], \] whose asymptotic variance is \[ Var(\hat{\theta}) = \frac{1}{n}\,Var(Y) + \frac{1}{n}\,E\!\left[ r_{I(X)}(X)\!\left(\frac{1}{\pi_{I(X)}(X)} - 1\right) \right]. \] The optimization problem is \[ \min_{\pi,\, \lambda,\, I} \; E\!\left[ r_{I(X)}(X)\!\left(\frac{1}{\pi_{I(X)}(X)} - 1\right) \right] \quad \text{s.t.} \quad E\!\left[ c_{I(X)} + \pi_{I(X)}(X) \cdot c_{\mathrm{label}} \right] \leq b, \] and the KKT conditions yield the proportional-to-residual-uncertainty rule \[ \pi^*_I(x) = \sqrt{\frac{r_I(x)}{n\,\mu \, c_{\text{label}}}}, \] clipped at $1$, together with weighted least-squares reweighting and per-instance routing [2605.08429]. 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 $10$ to $40$ percent narrower confidence intervals than single-predictor ASI in the routing-relevant regime [2605.08429]. Budget-constrained acquisition can also trade ground-truth labels against pairwise preferences. PCAL casts this as monotone missingness with observation patterns $r_1^*,r_2^*,r_3^*$ and budget constraint \[ E[ \rho \alpha_1(X, W_1, W_2) + \alpha_2(X, W_1, W_2) ] \le \tau. \] Its asymptotic variance objective is minimized over acquisition policies $\alpha$, and under MCAR the paper states that if \[ \rho > \frac{e_1 - e_3}{e_2 - e_3}, \] then the variance-minimizing allocation uses a nonzero $\alpha_2$, meaning preferences should be purchased [2601.13458]. The estimator is asymptotically normal, semiparametrically efficient for the chosen policy, and enjoys the robustness guarantee \[ \text{PCAL} \le \text{label-unlabeled} \le \text{label-only} \] in trace-variance, asymptotically, up to small tolerances [2601.13458]. In randomized experiments, AI-generated predictions enter as prognostic covariates rather than replacement outcomes. The recommended estimator is Lin-style regression adjustment, \[ Y_i = \alpha + \tau T_i + \beta^\top \tilde X_i + \gamma^\top \big(T_i \cdot \tilde X_i\big) + \varepsilon_i, \] with AI scores included in $X_i$ [2606.08853]. Under complete randomization, the difference-in-means variance is \[ \operatorname{Var}(\hat\tau_{DIM}) \;=\; \frac{S_1^2}{n_1} \;+\; \frac{S_0^2}{n_0} \;-\; \frac{S_\tau^2}{n}, \] and regression adjustment reduces the within-arm component by the prognostic correlation, with arm-specific design effect \[ \text{Deff}_z = 1 - \rho(z,z)^2. \] The paper emphasizes a “do no harm” property: if AI scores are uninformative, $\hat\tau_{RA}$ reverts to $\hat\tau_{DIM}$, whereas an uncalibrated model-assisted estimator has $\text{Deff}_z = 2(1-\rho(z,z))$ and inflates variance whenever $\rho(z,z)\le 0.5$ [2606.08853]. ## 5. Compute-time allocation, dynamic censoring, and generative dispersion Variance budgeting also appears in reinforcement learning for anytime reasoning. AnytimeReasoner defines \[ J_{\text{anytime}}(\theta,\phi) = E_{x,z}\!\left[\sum_{j=1}^{B} P_j \cdot r_\phi(x, z_{\le b_j})\right], \] where budgets $b_j$ are sampled from a prior and dense verifiable rewards are produced by summarizing truncated reasoning traces [2505.13438]. For the thinking policy, BRPO replaces the GRPO group-average baseline with an interpolation \[ V(x, z_{