Balanced Iteration Subsampling (BIS)

• Balanced Iteration Subsampling (BIS) is a structured subsampling method that enforces balance across iterations, sources, or sample participation to improve efficiency across domains.
• BIS is applied in multimodal reward models, imbalanced semi-supervised learning, block models, and DP-SGD, each using tailored selection strategies to boost performance and stability.
• By replacing naive random thinning with engineered selection, BIS achieves faster convergence, balanced classifiers, and stronger privacy guarantees in DP-SGD.

Balanced Iteration Subsampling (BIS) is not a single canonical algorithm but a recurring label for structured subsampling and re-sampling schemes that impose balance across iterations, sources, or sample participation. In the literature considered here, the acronym covers a rollout-selection score for Multimodal Process Reward Models, a progressive bi-sampling schedule for long-tailed semi-supervised learning, a subsampled power-iteration method for stochastic block models and planted CSPs, and a fixed-count participation rule for DP-SGD (Li et al., 4 Feb 2026, He et al., 2021, Feldman et al., 2014, Dong et al., 8 May 2026). Across these uses, BIS denotes a departure from naive random thinning: selection is engineered to preserve informative uncertainty, improve classifier balance, stabilize iterative signal extraction, or eliminate participation-count variance.

1. Terminological scope and cross-domain usage

The term BIS is polysemous in current research usage. In multimodal reward modeling, the acronym refers to the Balanced-Information Score rather than to “Balanced Iteration Subsampling.” In imbalanced semi-supervised learning, the corresponding method is presented as Bi-Sampling (BiS). In theoretical work on block models and planted CSPs, BIS appears as an exposition of Subsampled Power Iteration. In differentially private optimization, BIS denotes a specific Balanced Iteration Subsampling scheme for DP-SGD (Li et al., 4 Feb 2026, He et al., 2021, Feldman et al., 2014, Dong et al., 8 May 2026).

Setting	BIS formulation	Core operation
Multimodal PRMs	Balanced-Information Score	Rank rollouts by mixture and reliability from stored MC scores
Imbalanced SSL	Bi-Sampling (BiS)	Interpolate between two class samplers during training
Block models / planted CSPs	Subsampled Power Iteration	Split edges into subsamples and alternate centered matrix multiplies
DP-SGD	Balanced Iteration Subsampling	Force each sample to appear in exactly a fixed number of iterations

This breadth of usage matters because the shared acronym can obscure substantial methodological differences. What is common is the use of structured selection to modify the statistical properties of training or inference, but the objective function and operating regime differ sharply across domains.

2. BIS as Balanced-Information Score in multimodal process reward modeling

In Multimodal Process Reward Models (MPRMs), BIS is introduced to improve training-data efficiency under MC-annotated supervision. The motivating framework uses a teacher–student abstraction in which each reasoning step is represented by a feature vector $\phi \in \mathbb{R}^d$ with true binary label $Y \in \{0,1\}$, teacher output $q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$, and student model $q_w(\phi)=\sigma(w^\top\phi)$. At the teacher optimum, the squared gradient norm satisfies

$$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$

so informativeness is highest when $q^*(\phi)\approx \tfrac12$. Under label-flip probability $\eta$, the noisy-label gradient variance becomes

$$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$

which directly links training value to both label mixture and label reliability. Rollout-level uncertainty is proxied by $\hat p_x(1-\hat p_x)$, where $\hat p_x=(1/n)\sum Y_j$, and MC scores act as reliability indicators because among positive steps $Y \in \{0,1\}$0 the probability that $Y \in \{0,1\}$1 is a decreasing function of $Y \in \{0,1\}$2. BIS therefore combines mixture and reliability in an offline score computed from existing MC signals (Li et al., 4 Feb 2026).

For a rollout $Y \in \{0,1\}$3 with MC scores $Y \in \{0,1\}$4 and hard labels $Y \in \{0,1\}$5, the score is defined from the positive ratio

$Y \in \{0,1\}$6

the mixture proxy

$Y \in \{0,1\}$7

and the reliability proxy

$Y \in \{0,1\}$8

With smoothing constant $Y \in \{0,1\}$9, using $q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$0 throughout, the score is

$q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$1

Subsampling is single-pass and source-wise: compute MC scores and hard labels, evaluate $q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$2, $q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$3, $q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$4, and $q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$5, sort rollouts in descending BIS, and keep the top $q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$6 rollouts. No model calls or re-annotation are required; the method uses only stored MC scores. For scalability, BIS can be calculated in one streaming pass with per-source partial-sort to top-$q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$7 in $q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$8 time.

Empirically, on VisualPRM400K-v1.1 training and VisualProcessBench testing, BIS-selected subsets consistently match and even surpass the full-data performance at small fractions across InternVL2.5-8B and Qwen2.5-VL-7B. For InternVL2.5-8B, Full-Data reaches 65.12% overall micro-F1, Random-10% reaches 62.86%, and BIS-10% reaches 65.46%. For Qwen2.5-VL-7B, BIS-5% reaches 64.42% versus Random-5% at 53.54%, with Full at 65.57%. Figure 1 reports that for InternVL2.5-8B the BIS-10% performance curve converges in only 50 steps versus 1 100 for Full. At $q^*(\phi)=\Pr[Y=1\mid \phi]=\sigma(w^{*\top}\phi)$9, BIS-25% attains 65.46%, compared with 64.70% for Mixed-25%, 64.18% for Low-MC-25%, 62.75% for Reliable-25%, and 63.37% for Random-25%. The paper also reports stable performance across $q_w(\phi)=\sigma(w^\top\phi)$0 with the best result at $q_w(\phi)=\sigma(w^\top\phi)$1, and notes that BIS ignores the current student model, uses only rollout-level aggregates, does not adapt over training, and suffers higher mixture noise for low-$q_w(\phi)=\sigma(w^\top\phi)$2 rollouts because of the $q_w(\phi)=\sigma(w^\top\phi)$3 bias.

3. Bi-Sampling in imbalanced semi-supervised learning

In imbalanced semi-supervised learning, Bi-Sampling addresses the mismatch between representation learning and classifier learning under long-tailed class distributions. The central empirical observation is twofold: random sampling yields biased pseudo-labels dominated by head classes, yet those biased pseudo-labels do not harm representation learning and in fact yield the best feature extractors when the classifier is later re-balanced; conversely, re-sampling helps train a less-biased classifier but hurts representation quality if used throughout. Bi-Sampling therefore maintains two samplers in a single end-to-end run: Sampler A, such as pure random or no re-sampling, for learning the feature extractor, and Sampler B, such as mean or reverse re-sampling, for learning a balanced classifier (He et al., 2021).

The mechanism is a progressive interpolation. If $q_w(\phi)=\sigma(w^\top\phi)$4 is the labeled count for class $q_w(\phi)=\sigma(w^\top\phi)$5, $q_w(\phi)=\sigma(w^\top\phi)$6, and $q_w(\phi)=\sigma(w^\top\phi)$7 is the number of classes, then Sampler A uses

$q_w(\phi)=\sigma(w^\top\phi)$8

while a balanced Sampler B uses

$q_w(\phi)=\sigma(w^\top\phi)$9

or, in a reverse-sampling variant,

$$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$0

At epoch $$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$1, the combined class weight is

$$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$2

with parabolic decay

$$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$3

For unlabeled pseudo-labels, the keep probability is

$$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$4

The design ensures that early training is dominated by the natural imbalanced distribution, while late training is increasingly re-balanced for classifier reliability. The paper reports that parabolic decay outperforms linear, cosine, and static $$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$5 schedules.

The experimental evidence emphasizes minority-class recall. On CIFAR-10 with imbalance $$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$6 and unlabeled ratio $$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$7, Class 10 recall increases from 37.2% for FixMatch to 68.6% for FixMatch+BiS, and Class 9 recall increases from 46.8% to 69.9%, while head-class degradation is limited, with Class 1 moving from 98.3% to 94.4%. Overall accuracy increases from 73.9% to 81.0%. On CIFAR-10 with $$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$8 and $$E_{(Y\mid \phi)}[\|\nabla \ell(w^*;\phi,Y)\|^2] = q^*(\phi)[1-q^*(\phi)]\cdot\|\phi\|^2,$$9 in the FixMatch family, the reported accuracies are 62.4% for random baseline, 68.3% for mean sampling, 68.9% for reverse sampling, 66.2% for DARP+cRT, 69.8% for CReST+, and 71.0% for BiS. The associated misconception addressed by this work is explicit: accurate pseudo-labels do not necessarily help feature extractor training, and data re-sampling can harm representation learning even when it improves classifier balance.

4. Subsampled power iteration for block models and planted CSPs

In the stochastic block model and planted CSP setting, the BIS formulation is a subsampled power-iteration procedure over a random bipartite graph. The underlying model is

$q^*(\phi)\approx \tfrac12$0

where $q^*(\phi)\approx \tfrac12$1 have sizes $q^*(\phi)\approx \tfrac12$2, each partitioned into $q^*(\phi)\approx \tfrac12$3 and $q^*(\phi)\approx \tfrac12$4, and edges between matching parts appear with probability $q^*(\phi)\approx \tfrac12$5 while cross edges appear with probability $q^*(\phi)\approx \tfrac12$6. The edge set is split into $q^*(\phi)\approx \tfrac12$7 roughly equal, independent subsamples, with $q^*(\phi)\approx \tfrac12$8 and the concrete choice $q^*(\phi)\approx \tfrac12$9. For each bin $\eta$0, if $\eta$1 is the bipartite adjacency matrix of that subsample, the centered matrix is

$\eta$2

where $\eta$3 is the all-ones $\eta$4 matrix. Starting from a random unit vector $\eta$5, the algorithm alternates

$\eta$6

with normalization after each multiplication, records $\eta$7, and outputs a coordinate-wise majority vote over the last $\eta$8 iterates (Feldman et al., 2014).

The central theorem states that there is an absolute constant $\eta$9 such that if

$$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$0

then the algorithm with $$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$1 recovers the true partition $$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$2 on $$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$3 exactly with probability $$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$4 as $$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$5. Its running time is

$$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$6

and its space usage is $$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$7. In the planted CSP reduction, the same approach yields recovery of the planted assignment in linear time under the stated density scaling. The analysis tracks the potentials $$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$8 and $$E[\|\tilde g(\phi;w^*)\|^2] = [(1-4\eta)\cdot q^*(\phi)(1-q^*(\phi))+\eta]\cdot \|\phi\|^2,$$9, proceeds through initial growth, geometric growth, constant correlation, and majority-vote phases, and requires concentration bounds for expectations, variances, normalizers, and $\hat p_x(1-\hat p_x)$0 control.

The significance of subsampling here differs from its role in learning pipelines. The point is not class balance or privacy but the creation of nearly independent centered operators that allow the iteration to outperform spectral approaches in highly unbalanced regimes. The comparison given in the exposition is explicit: spectral methods require $\hat p_x(1-\hat p_x)$1 in the regime $\hat p_x(1-\hat p_x)$2, whereas BIS succeeds at $\hat p_x(1-\hat p_x)$3; equivalently, spectral methods need $\hat p_x(1-\hat p_x)$4 while BIS needs only $\hat p_x(1-\hat p_x)$5.

5. BIS in DP-SGD and privacy amplification

In DP-SGD, BIS is a structured subsampling scheme designed to eliminate participation variance while keeping uniform marginal participation across iterations. Let $\hat p_x(1-\hat p_x)$6 be the dataset size, $\hat p_x(1-\hat p_x)$7 the number of training iterations, $\hat p_x(1-\hat p_x)$8 the batch size, and

$\hat p_x(1-\hat p_x)$9

with the assumption that $\hat p_x=(1/n)\sum Y_j$0. Under BIS, each example $\hat p_x=(1/n)\sum Y_j$1 is included in exactly $\hat p_x=(1/n)\sum Y_j$2 of the $\hat p_x=(1/n)\sum Y_j$3 iterations: one draws

$\hat p_x=(1/n)\sum Y_j$4

uniformly at random among all subsets of size $\hat p_x=(1/n)\sum Y_j$5, and defines the minibatch at iteration $\hat p_x=(1/n)\sum Y_j$6 by

$\hat p_x=(1/n)\sum Y_j$7

Equivalently, each example’s participation vector is drawn i.i.d. from the unique distribution on $\hat p_x=(1/n)\sum Y_j$8 satisfying $\hat p_x=(1/n)\sum Y_j$9 with probability $Y \in \{0,1\}$00 and assigning probability $Y \in \{0,1\}$01 to each support pattern of Hamming weight $Y \in \{0,1\}$02. The DP-SGD update then clips gradients for $Y \in \{0,1\}$03, averages them, adds Gaussian noise $Y \in \{0,1\}$04, and updates parameters. This contrasts with Poisson subsampling, in which each example is included independently at each iteration with probability $Y \in \{0,1\}$05 (Dong et al., 8 May 2026).

The privacy analysis is formulated under zero-out adjacency. In the low-noise regime $Y \in \{0,1\}$06, if $Y \in \{0,1\}$07 with $Y \in \{0,1\}$08 and $Y \in \{0,1\}$09, the privacy loss has leading term proportional to $Y \in \{0,1\}$10. Under BIS, $Y \in \{0,1\}$11 is nonrandom, which eliminates the dominant source of privacy-loss tail that appears under Poisson subsampling because $Y \in \{0,1\}$12. The corresponding theorem states that among all independent-example subsampling schemes with $Y \in \{0,1\}$13, BIS uniquely minimizes the upper tail of the privacy loss as $Y \in \{0,1\}$14. In the high-noise regime $Y \in \{0,1\}$15, the first-order term depends only on uniform marginals, while the second-order term is governed by the covariance matrix $Y \in \{0,1\}$16; BIS fixes $Y \in \{0,1\}$17, so $Y \in \{0,1\}$18, and by symmetry distributes the remaining variance isotropically on $Y \in \{0,1\}$19, yielding the best second-order constant. The paper therefore states optimality at both extremes of the noise spectrum.

For finite-noise accounting, the work introduces a near-exact Monte Carlo accountant. Exact $Y \in \{0,1\}$20-DP requires

$Y \in \{0,1\}$21

Under BIS,

$Y \in \{0,1\}$22

and

$Y \in \{0,1\}$23

With

$Y \in \{0,1\}$24

the likelihood ratio can be evaluated by dynamic programming through the recurrence

$Y \in \{0,1\}$25

with $Y \in \{0,1\}$26 time and $Y \in \{0,1\}$27 memory per sample. A screening bound provides an $Y \in \{0,1\}$28 shortcut for samples that cannot contribute to $Y \in \{0,1\}$29. Empirically, across more than 60 practical DP-SGD configurations, BIS-MC requires up to 9.6% less noise multiplier than Poisson in low-noise regimes, with average reduction approximately 5–8%, while in high-noise regimes the required noise converges to Poisson-PLD within at most 1%. The stated practical conclusion is that BIS is a drop-in replacement for Poisson subsampling with the same expected batch size and gradient-estimate variance but stronger privacy amplification.

6. Common themes, distinctions, and recurrent misconceptions

A common thread across these formulations is that subsampling is treated as a statistical design variable rather than as incidental randomness. In MPRM training, BIS uses stored MC annotations to favor rollouts with both uncertainty and reliability instead of relying on random subsampling (Li et al., 4 Feb 2026). In long-tailed SSL, BiS progressively shifts from the natural imbalanced distribution to a balanced classifier-oriented distribution rather than adopting a single fixed re-sampling rule (He et al., 2021). In block models and planted CSPs, subsampling creates a sequence of centered operators for alternating power iteration, which is analytically and computationally preferable to operating on the full matrix in unbalanced regimes (Feldman et al., 2014). In DP-SGD, BIS enforces exact participation counts and thereby removes the variance that Poisson subsampling leaves uncontrolled (Dong et al., 8 May 2026).

Several misconceptions are directly contradicted by these results. More data is not always more useful under fixed annotation structure: in MPRM training, random subsampling quickly saturates, and the BIS subset reaches full-data performance using only 10% of the training data. More accurate or more balanced pseudo-labels are not universally beneficial during SSL: the cited BiS study finds that re-sampling can hurt the feature extractor even while helping the classifier. More sampling randomness is not necessarily better for privacy amplification: the DP-SGD BIS paper argues the opposite in the low-noise regime, where participation-count variance is the dominant source of privacy-loss tail. Even the acronym itself invites confusion, since “BIS” may denote Balanced-Information Score, Bi-Sampling, a subsampled power-iteration framework, or Balanced Iteration Subsampling proper.

This suggests a broader interpretation of BIS as a family of structured-selection principles. The precise balance being enforced differs by field—positive/negative label mixture and MC reliability, head/tail class frequencies over training time, near-independent edge partitions for iterative recovery, or fixed sample participation counts—but the unifying idea is the same: replacing uncontrolled randomness with constrained sampling structure can improve efficiency, recovery, or privacy under the objective relevant to the domain.