Boundary-Aware Exemplar Prioritization

• Boundary-Aware Exemplar Prioritization is a design principle that selects informative examples based on their proximity to task-specific decision boundaries rather than traditional relevance metrics.
• It employs asymmetric scoring functions—such as modified ambiguity sampling and enemy-aware synthesis—to target useful positives while mitigating redundancy and instability.
• Applications span rare-category retrieval, continual learning, and retrieval-augmented generation, enhancing data utility and model refinement under challenging distributions.

Boundary-aware exemplar prioritization denotes a family of selection, retention, or synthesis strategies in which examples are valued according to their relation to a task-relevant boundary rather than by relevance, density, or uniform sampling alone. In the recent literature, that boundary is instantiated in several distinct ways: a classifier decision threshold in interactive rare-category retrieval, an uncertainty or confidence boundary in preference alignment, an old/new class frontier in exemplar-free continual learning, a generator competence boundary in retrieval-augmented generation, and, more loosely, sparse semantic frontiers or image-edge-aligned uncertainty in corpus curation and representation learning (Zaher et al., 25 Mar 2026, Wu et al., 1 Feb 2026, Xu et al., 4 Jun 2026, Sun et al., 3 Feb 2026, Wu et al., 27 May 2026, Peng et al., 2022, Wåhlstrand et al., 16 Mar 2026). Taken together, these works suggest that the topic is best understood as a design principle for data selection under asymmetry, redundancy, or shifting utility, rather than as a single algorithmic template.

1. Conceptual scope and recurrent boundary notions

The most direct formulations treat exemplar prioritization as a ranking problem over candidate items. In "Positive-First Most Ambiguous" (Zaher et al., 25 Mar 2026), the ranked items are unlabeled images in an interactive retrieval loop. In SAGE (Wu et al., 1 Feb 2026), they are preference pairs in reasoning-model alignment. In "Revisiting Prototype Rehearsal for Exemplar-Free Continual Learning" (Xu et al., 4 Jun 2026), they are synthetic old-class features generated from prototypes and enemy classes. In BAR-RAG (Sun et al., 3 Feb 2026), they are evidence sets for a generator. GUI-CIDER instead performs stochastic retention over synthesized textual GUI knowledge, while "Boundary-aware Information Maximization for Self-supervised Medical Image Segmentation" and "Exemplar Diffusion" provide only indirect or implicit analogues (Wu et al., 27 May 2026, Peng et al., 2022, Wåhlstrand et al., 16 Mar 2026).

Work	Prioritized unit	Boundary notion
PF-MA	Unlabeled retrieval candidates	$f(x)=0.5$ decision threshold
SAGE	Preference pairs	$p \approx 0.5$ confidence boundary
CEOS + ACB	Synthetic old-class features	Old/new frontier via nearest enemy
BAR-RAG	Evidence sets	Goldilocks Zone $p(S)\approx c$
GUI-CIDER	Synthesized GUI knowledge samples	Sparse semantic regions and dense-region rescue
Boundary-aware Information Maximization	Pixels or regions implicitly	Entropy transitions aligned to edges
Exemplar Diffusion	Inference-time boxes implicitly	Spatial anchor and extent prior

A recurrent contrast in these papers is between three regimes. One regime consists of easy or dense-core examples, which are often redundant. A second consists of maximally ambiguous or unstable examples, which can be noisy or utility-mismatched. A third consists of boundary-adjacent but still useful examples: informative positives in PF-MA, confident errors in SAGE, nearest-enemy-constrained synthetic old features in CEOS, or evidence sets near a generator’s competence boundary in BAR-RAG. This suggests that “boundary-aware” does not mean indiscriminate preference for the most ambiguous samples; rather, it typically means selective exploitation of a frontier region defined by task utility.

2. Positive-side boundary sampling in interactive rare-category retrieval

The clearest explicit formulation appears in "Positive-First Most Ambiguous: A Simple Active Learning Criterion for Interactive Retrieval of Rare Categories" (Zaher et al., 25 Mar 2026). The problem is an imbalanced binary retrieval task over a large image collection, with positive class equal to the user’s class-of-interest and negative class equal to everything else. Images are embedded as

$D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$

The labeled set is initialized from a tiny query with $N_p=1$ positive and $N_n=5$ negatives; the system runs for up to $T=25$ rounds with batch budget $b=10$. The paper frames this as severe class imbalance, tiny per-iteration annotation budgets, and low latency, with the objective of maximizing useful rare-class discovery early rather than asymptotic classifier accuracy.

The paper’s critique of conventional active learning is that standard ambiguity sampling assumes roughly symmetric class priors, symmetric labeling utility, and enough budget for uncertainty-driven exploration to pay off. Under rare-class frequencies such as $1/1000$ or $1/10000$, symmetric boundary sampling mostly returns negatives. The baseline "Most Ambiguous" score,

$p \approx 0.5$0

is boundary-aware but oversamples near-boundary negatives. The baseline "Most Positive" score,

$p \approx 0.5$1

returns likely positives quickly but tends to retrieve easy and redundant positives.

PF-MA is the asymmetric alternative: $p \approx 0.5$2 On the positive half-space, the score is ambiguity-based; on the negative half-space, it is a positive-likelihood score. The paper states that PF-MA first selects positives near the decision boundary, then less ambiguous positives, and only includes negatives if the iteration budget is not yet filled. The classifier is a linear SVM producing scores $p \approx 0.5$3, and the method is designed for frozen pretrained descriptors, specifically CLIP and DINOv2.

The paper also introduces a class coverage metric to quantify whether retrieved positives span the internal variability of the target class. For class $p \approx 0.5$4 with retrieved positives $p \approx 0.5$5, coverage is defined by clustering class instances with K-means and measuring the fraction of clusters hit: $p \approx 0.5$6 with main choice $p \approx 0.5$7, averaged over 10 K-means runs.

Empirically, PF-MA is best or tied-best nearly everywhere in class coverage at iteration 25. On ImageNet-LT with DINOv2, for example, $p \approx 0.5$8 is $p \approx 0.5$9 for MA, $p(S)\approx c$0 for MP, and $p(S)\approx c$1 for PF-MA; $p(S)\approx c$2 is $p(S)\approx c$3, $p(S)\approx c$4, and $p(S)\approx c$5; and $p(S)\approx c$6 is $p(S)\approx c$7, $p(S)\approx c$8, and $p(S)\approx c$9 (Zaher et al., 25 Mar 2026). On CIFAR100-LT and ImageNet-LT, PF-MA keeps more than $D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$0 positives in the selected batch with low variance across iterations while remaining top-tier in F1. The paper’s interpretation is that these are not merely easy positives, but boundary-adjacent positives that remain informative for classifier refinement.

PF-MA also clarifies an important misconception. Boundary awareness here is not raw SVM hyperplane-distance sampling, nor ordinary symmetric uncertainty sampling. It is an asymmetric rule that preserves boundary informativeness while restricting the primary search region to the predicted-positive side of the decision boundary. In that sense, it is exemplar prioritization under label asymmetry.

3. Confidence-boundary filtering in preference alignment

SAGE, introduced in "Not All Preferences Are Created Equal: Stability-Aware and Gradient-Efficient Alignment for Reasoning Models" (Wu et al., 1 Feb 2026), recasts exemplar prioritization for preference pairs. The paper argues that standard DPO-style alignment treats all preference pairs uniformly even though sample utility depends on the current policy state $D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$1. Two failure modes are emphasized: trivial pairs with negligible gradients, and unstable pairs near uncertain decision boundaries, where the local loss geometry is high-curvature and updates can be noisy.

The method therefore defines a dynamic, policy-aware prioritization scheme with two layers. The coarse layer is a curriculum over disjoint candidate pools

$D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$2

sampled from easy, medium, and hard strata according to a linearly evolving mixture $D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$3. The appendix gives the concrete endpoints: start mix $D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$4, end mix $D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$5, and keep-ratio schedule from $D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$6 to $D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$7.

The fine layer uses a stability-aware score. For the response-level logistic term $D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$8, with $D=\{x_i\}_{i\in[1,N]}, \qquad x_i=\Phi(I_i)\in\mathds{R}^d.$9, the curvature proxy is

$N_p=1$0

and the gradient-signal proxy satisfies

$N_p=1$1

The resulting sample score is

$N_p=1$2

Because $N_p=1$3 peaks at $N_p=1$4, the score explicitly suppresses samples near the model’s uncertainty boundary. The preferred exemplars are therefore informative, confident errors rather than maximally ambiguous pairs.

This is a boundary-aware formulation in a negative sense: the boundary is treated as a warning signal rather than a target region. The paper is explicit that the method does not prioritize near-boundary examples indiscriminately. Instead, it rejects already-solved points with tiny $N_p=1$5, rejects highly uncertain points with large $N_p=1$6, and favors examples with strong residual error but lower instability. That distinction is central to the topic: boundary-aware exemplar prioritization can either seek a boundary region or avoid it, depending on how utility is defined.

The empirical evidence is consistent with this view. Across Qwen2.5-Instruct 1.5B, 3B, and 7B, SAGE improves average accuracy from $N_p=1$7 to $N_p=1$8, $N_p=1$9 to $N_n=5$0, and $N_n=5$1 to $N_n=5$2, respectively. On 7B, Minerva improves from $N_n=5$3 to $N_n=5$4, AIME24 from $N_n=5$5 to $N_n=5$6, and AMC23 from $N_n=5$7 to $N_n=5$8. Ablations show that removing the curvature term $N_n=5$9 causes among the largest drops, and gradient-norm plots show lower scale and lower variance than standard DPO (Wu et al., 1 Feb 2026). The paper interprets this as evidence that ambiguous preference pairs near uncertain confidence boundaries are actively harmful for optimization stability.

4. Nearest-enemy boundary synthesis in exemplar-free continual learning

"Revisiting Prototype Rehearsal for Exemplar-Free Continual Learning: Manifold-Aware Boundary Sampling with Adaptive Class-Balanced Loss" (Xu et al., 4 Jun 2026) shifts the topic from ranking existing items to generating synthetic old-class exemplars near old/new class frontiers. In exemplar-free class-incremental learning, old classes are summarized by prototypes

$T=25$0

and raw old samples cannot be stored. The paper argues that standard prototype rehearsal fails because it samples around isolated class centers, ignores nearby enemy classes, suffers from prototype drift, and induces a hidden old/new imbalance.

The proposed Constrained Expansive Over-Sampling (CEOS) uses current-task enemy features to construct boundary-aware synthetic old-class features. For a prototype $T=25$1 and its nearest enemy $T=25$2, the synthetic feature is

$T=25$3

The lower bound $T=25$4 enforces prototype dominance. The paper also defines a Mahalanobis metric

$T=25$5

and requires

$T=25$6

so that the synthetic sample remains on the prototype’s side of the local boundary. In experiments, CEOS selects one enemy per prototype, $T=25$7, which the paper reports as the best setting.

This construction makes the prioritized region explicit: not the prototype core, but the segment from old-class prototype toward the nearest confusing new-class feature, truncated so that the sample does not cross the boundary. The method is therefore boundary-aware, enemy-aware, and manifold-aware in the paper’s terminology. The paper’s asymptotic bias argument further motivates this design by claiming that with only $T=25$8 replay instances per old class, the expected old-class posterior under softmax vanishes asymptotically: $T=25$9

Boundary-aware generation is coupled with optimization-time weighting through the Adaptive Class-Balanced loss. For class $b=10$0 first seen at task $b=10$1, the virtual sample count is

$b=10$2

and the class-balanced weight is

$b=10$3

The paper’s interpretation is that freshly created prototypes should be amplified early and gradually attenuated as richer supervision accumulates. This is a temporal prioritization mechanism layered on top of geometric boundary targeting.

The ablations on TinyImageNet are especially direct. For $b=10$4, the base EFC-style baseline yields $b=10$5 and $b=10$6; CEOS only gives $b=10$7 and $b=10$8; ACB only gives $b=10$9 and $1/1000$0; and CEOS + ACB gives $1/1000$1 and $1/1000$2. Table 4 further shows that increasing Gaussian prototype samples from $1/1000$3 to $1/1000$4 or $1/1000$5 does not help, while CEOS at $1/1000$6 is better than Gaussian and bi-interpolation on both $1/1000$7 and $1/1000$8 (Xu et al., 4 Jun 2026). The paper’s conclusion is that the decisive factor is where synthetic replay support is placed in feature space, not merely how many synthetic samples are produced.

5. Goldilocks and density frontiers in evidence and corpus selection

Two 2026 papers broaden the topic beyond geometric class boundaries. "Rethinking the Reranker: Boundary-Aware Evidence Selection for Robust Retrieval-Augmented Generation" (Sun et al., 3 Feb 2026) defines the relevant boundary in terms of generator competence. For an evidence set $1/1000$9, let $1/10000$0 be the empirical probability that the generator answers correctly. BAR-RAG seeks the generator’s Goldilocks Zone, where evidence is neither trivially easy nor fundamentally unanswerable. The core boundary reward is

$1/10000$1

typically with $1/10000$2. Evidence with $1/10000$3 is too easy; evidence with $1/10000$4 is too hard; evidence with $1/10000$5 is challenging yet sufficient.

The selector is trained by reinforcement learning with generator feedback. Candidate evidence sets are sampled from the selector, the frozen generator is rolled out multiple times to estimate $1/10000$6, and the selector reward combines the boundary term, a relevance term, a format-validity gate, and a count penalty. The paper also filters training queries whose mean correctness lies outside $1/10000$7 or whose variance is too low, specifically using $1/10000$8, $1/10000$9, and $p \approx 0.5$00, so that evidence selection actually matters during selector learning. This yields a dynamic competence-aware notion of exemplar utility.

Empirically, BAR-RAG reports an average gain of $p \approx 0.5$01 percent over strong baselines. Average EM improves from $p \approx 0.5$02 to $p \approx 0.5$03 for Qwen-2.5-3B-Instruct, from $p \approx 0.5$04 to $p \approx 0.5$05 for Qwen-2.5-7B-Instruct, and from $p \approx 0.5$06 to $p \approx 0.5$07 for LLaMA-3.1-8B-Instruct. The ablation without $p \approx 0.5$08 drops average performance from $p \approx 0.5$09 to $p \approx 0.5$10, which the paper treats as evidence that boundary awareness, rather than relevance alone, is the key ingredient (Sun et al., 3 Feb 2026).

GUI-CIDER, by contrast, is not explicitly boundary-aware in the classification-margin sense. "GUI-CIDER: Mid-training GUI Agents via Causal Internalization and Density-aware Exemplar Reselection" (Wu et al., 27 May 2026) defines a density-aware, causality-weighted retention rule over synthesized textual GUI knowledge. The causal saliency score is

$p \approx 0.5$11

the normalized density score is $p \approx 0.5$12, and the retention probability is

$p \approx 0.5$13

The method performs stochastic thinning rather than deterministic top-$p \approx 0.5$14 ranking: $p \approx 0.5$15

The theoretical properties are explicit. $p \approx 0.5$16 increases with $p \approx 0.5$17, decreases with $p \approx 0.5$18, satisfies $p \approx 0.5$19, and has positive mixed derivative

$p \approx 0.5$20

The paper interprets this as making “the marginal value of causal saliency” larger in denser regions, so that “the most logically rigorous exemplars” stand out among redundant samples. Strictly speaking, this is not decision-boundary prioritization. A plausible implication is that it acts as a weak frontier-preservation mechanism by favoring sparse semantic regions and rescuing decision-relevant samples inside dense clusters. The Stage 2 ablation supports the importance of this reselection: on GUI-Odyssey, Qwen3-VL-4B-Instruct improves from $p \approx 0.5$21 to $p \approx 0.5$22, and Qwen3-VL-8B-Instruct from $p \approx 0.5$23 to $p \approx 0.5$24, when Stage 2 is included (Wu et al., 27 May 2026).

These two papers illustrate a broader shift. In BAR-RAG, the useful boundary is a competence boundary in downstream generation. In GUI-CIDER, the useful frontier is a semantic-density boundary modulated by causal structure. In both cases, exemplar value is defined relative to the needs of a downstream model rather than by relevance or representativeness alone.

6. Implicit boundary emphasis, misconceptions, and open issues

Some relevant works are adjacent to the topic without providing explicit exemplar prioritization. "Boundary-aware Information Maximization for Self-supervised Medical Image Segmentation" (Peng et al., 2022) is a case in point. The paper learns pixelwise cluster distributions over $p \approx 0.5$25 clusters and uses the entropy of the cluster distribution as a boundary proxy: $p \approx 0.5$26 High entropy indicates uncertainty between clusters and tends to occur near transitions. The boundary-aware term aligns this entropy map with Sobel edge responses $p \approx 0.5$27 using local normalized cross-correlation. The method therefore treats interior and boundary pixels differently—confident assignments are encouraged in interiors, while uncertainty is structurally permitted near edges—but it does not rank exemplars, perform top-$p \approx 0.5$28 selection, or implement hard mining. The paper explicitly states that it provides boundary-aware representation learning rather than a concrete exemplar prioritization mechanism. Its quantitative gains, however, show the practical importance of boundary-sensitive structure: the strongest pre-training configuration reports mean DSC improvements over baseline of $p \approx 0.5$29 for LV, $p \approx 0.5$30 for RV, $p \approx 0.5$31 for Myo, and $p \approx 0.5$32 for prostate (Peng et al., 2022).

"Exemplar Diffusion: Improving Medical Object Detection with Opportunistic Labels" (Wåhlstrand et al., 16 Mar 2026) is similarly indirect. Here the exemplar is a known bounding box at inference time, $p \approx 0.5$33, and the contribution is a training-free mechanism for injecting lightly noised exemplars into the proposal distribution of a DiffusionDet-style detector: $p \approx 0.5$34 This is spatially anchored conditioning rather than explicit boundary prioritization. The paper emphasizes that the method is most useful when images have clear spatial structure. On DENTEX, AP improves from $p \approx 0.5$35 with $p \approx 0.5$36 to $p \approx 0.5$37 with $p \approx 0.5$38 and $p \approx 0.5$39 with $p \approx 0.5$40; $p \approx 0.5$41 rises from $p \approx 0.5$42 to $p \approx 0.5$43 and $p \approx 0.5$44; and AR from $p \approx 0.5$45 to $p \approx 0.5$46 and $p \approx 0.5$47. The method is also robust to noisy exemplars: with one exemplar and Gaussian noise scale $p \approx 0.5$48, AP is $p \approx 0.5$49, still near baseline (Wåhlstrand et al., 16 Mar 2026). The paper’s own framing supports a boundary-aware interpretation only weakly: box coordinates provide a spatial anchor and extent prior, but the method does not rank exemplars by boundary informativeness.

These cases help clarify several misconceptions. First, boundary-aware prioritization is not synonymous with uncertainty sampling. PF-MA modifies ambiguity sampling to prefer the positive half-space, whereas SAGE explicitly suppresses samples near unstable uncertainty boundaries (Zaher et al., 25 Mar 2026, Wu et al., 1 Feb 2026). Second, the relevant boundary need not be geometric. BAR-RAG’s boundary is defined by empirical solvability for a generator, and GUI-CIDER’s frontier is defined by density and causal saliency (Sun et al., 3 Feb 2026, Wu et al., 27 May 2026). Third, some methods are boundary-aware only in an implicit structural sense. The medical segmentation work and exemplar diffusion improve boundary-sensitive behavior without implementing exemplar ranking at all (Peng et al., 2022, Wåhlstrand et al., 16 Mar 2026).

A plausible synthesis is that the field is converging on a utility-centric view of exemplar value. Under severe imbalance, scarce annotation, replay without stored data, noisy retrieval, or redundant synthetic corpora, the highest-value examples are often neither dense-core nor maximally ambiguous. They are examples that occupy a controlled frontier: informative positives near a threshold, confident errors away from unstable uncertainty, synthetic points approaching but not crossing a class boundary, evidence near a generator’s Goldilocks Zone, or sparse and causally rich knowledge units. That shared structure is what makes boundary-aware exemplar prioritization a coherent research topic across otherwise disparate areas.