Boundary-Aware Contrastive Sampling (BACS)

• The paper introduces BACS as a boundary-focused sampling method that transforms noisy batches into a compact set of high-fidelity examples near decision boundaries.
• Its methodology employs semantic clustering and error-rate guided quotas to extract tripartite features—Hard Negatives, Anchors, and Boundary Pairs—that sharpen optimization signals.
• Empirical results show that integrating BACS in C-MOP significantly reduces gradient variance and enhances model robustness compared to naive sampling approaches.

Searching arXiv for the cited BACS-related papers to ground the article in current arXiv work. Search query: "Boundary-Aware Contrastive Sampling C-MOP arXiv (Yan et al., 11 Feb 2026)" Boundary-Aware Contrastive Sampling (BACS) is a boundary-focused sampling strategy for contrastive optimization whose explicit formulation appears as the first core component of C-MOP, where it converts a large, noisy batch of prompt examples into a compact set of highly informative cases near the decision boundary between success and failure (Yan et al., 11 Feb 2026). In that formulation, BACS mines three feature categories—Hard Negatives, Anchors, and Boundary Pairs—by clustering batch embeddings, allocating more sampling quota to clusters with higher error rates, and extracting prototypical success, systematic failure, and minimal success/failure contrasts. Closely related mechanisms appear in semantic segmentation, unsupervised node clustering, recommendation, and time-series anomaly detection, where boundary-aware negative construction is instantiated through error-boundary pixels, virtual nodes near cluster boundaries, decision-boundary-constrained perturbations, or reconstruction-induced boundary shifts (Sung et al., 2024).

1. Definition and conceptual scope

Within C-MOP, BACS is defined as a batch-level sampling module that uses semantic clustering and error-aware quota allocation to mine tripartite features that characterize positive and negative prompt behavior (Yan et al., 11 Feb 2026). Its stated objectives are threefold: to capture the semantic centroids of correct and incorrect model behavior, to focus sampling on clusters with high error rates, and to isolate the minimal linguistic shifts that cause the model to flip its answer. The mechanism is explicitly designed for large-batch textual-gradient optimization, where simply increasing batch size brings in more examples but also more conflicting or uninformative gradients.

Formally, BACS begins from a sampled batch $D_t$ of size $B$. For each example $(q_i, a_i)$, the current prompt $p_t$ produces a prediction $r_i$, and a frozen encoder $\phi$ embeds the concatenated triplet: $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$ The $B$ embeddings are clustered into $K$ semantic clusters $C_1,\dots,C_K$. Each cluster is then partitioned into a positive set $B$0 of correct predictions and a negative set $B$1 of incorrect predictions, and an error rate is assigned as

$B$2

Sampling quota is normalized by these error rates, so clusters where the current prompt fails most receive more examples.

This construction places BACS within a broader family of boundary-aware hard-negative strategies, but the term is not used uniformly across the literature. In the graph-clustering paper "Unsupervised node clustering via contrastive hard sampling" (Cui et al., 2024), the source method is MeCoLe rather than a module literally named BACS; the provided exposition explicitly states that the paper does not introduce a module called “Boundary-Aware Contrastive Sampling” by that name. This suggests that BACS is best understood as a methodological pattern—boundary-focused contrastive sample construction—rather than a single canonical algorithm.

2. Tripartite feature mining and mathematical structure

The distinctive feature of BACS in C-MOP is its tripartite decomposition of sampled exemplars into Hard Negatives, Anchors, and Boundary Pairs (Yan et al., 11 Feb 2026). Let

$B$3

denote the centroid of cluster $B$4. The three feature types are then defined as follows.

A Hard Negative is the negative example closest to the centroid: $B$5 This selects a systematic failure rather than an outlier failure.

An Anchor is the prototypical correct example closest to the same centroid: $B$6 This serves as a stable success reference.

Boundary Pairs are the $B$7 positive-negative pairs that minimize cross-class distance within the cluster: $B$8 These pairs are intended to isolate the minimal semantic flip that toggles the model’s decision.

Quota allocation is defined by

$B$9

If the selected set $(q_i, a_i)$0 for cluster $(q_i, a_i)$1 exceeds $(q_i, a_i)$2, truncation prioritizes Boundary Pairs, then Hard Negatives, then Anchors. The priority ordering is consequential: the paper’s exposition assigns the highest value to minimal success/failure contrasts, then to systematic failures, and finally to prototypical successes.

The underlying contrastive intuition is explicit in the paper: contrastive learning is most effective when positive and negative examples lie close in representation space, that is, near the decision boundary. BACS therefore does not merely search for negative samples; it organizes local semantic geometry so that the optimizer observes representative successes, representative failures, and the narrow transition region between them (Yan et al., 11 Feb 2026).

3. Sampling procedure and integration within C-MOP

BACS operates as a subroutine inside each iteration of C-MOP’s optimization loop (Yan et al., 11 Feb 2026). The procedure can be summarized as follows. First, the current prompt $(q_i, a_i)$3 is evaluated on a sampled batch $(q_i, a_i)$4, producing predictions $(q_i, a_i)$5, and every triplet $(q_i, a_i)$6 is embedded using $(q_i, a_i)$7. Second, the embeddings are clustered via K-means into $(q_i, a_i)$8 semantic clusters. Third, cluster error rates $(q_i, a_i)$9 are computed and transformed into quotas $p_t$0. Fourth, each cluster yields a set of tripartite exemplars: one Hard Negative, one Anchor, and a small set of Boundary Pairs. Fifth, the union of these cluster-level selections forms the set $p_t$1, which is returned for gradient generation.

The default hyperparameters reported for BACS in C-MOP are: batch size $p_t$2, total sampling quota $p_t$3, number of clusters $p_t$4, number of boundary pairs per cluster $p_t$5, and embedding model all-MiniLM-L6-v2 (Yan et al., 11 Feb 2026). These values are part of the concrete implementation recipe given in the paper.

BACS is then coupled with Momentum-Guided Semantic Clustering (MGSC). In the main loop, BACS replaces ordinary stochastic or failure-only sampling in Step 10, and the selected set $p_t$6 is fed into the optimizer model to generate new textual gradients. Those gradients are subsequently merged with a decayed historical pool, clustered, and distilled into stable update directions by MGSC. The paper describes this as a tightly coupled two-stage refinement: BACS provides high-fidelity, boundary-focused raw signals, while MGSC aggregates and denoises them across iterations (Yan et al., 11 Feb 2026).

A plausible implication is that BACS alone is not intended to solve prompt optimization instability; rather, it reshapes the local sampling distribution so that downstream gradient aggregation has a cleaner signal to work with. That interpretation is consistent with the reported ablations in which BACS without MGSC improves over weaker sampling baselines but does not match the full framework.

4. Related boundary-aware contrastive mechanisms in other domains

Boundary-aware contrastive sampling has close analogues in several other problem settings, but the operational meaning of “boundary” differs across tasks.

Paper	Domain	Boundary-aware mechanism
"Contextrast: Contextual Contrastive Learning for Semantic Segmentation" (Sung et al., 2024)	Semantic segmentation	Error-boundary pixels are extracted from class-wise false-negative regions, ranked by distance transform, and the lowest-$p_t$7-percentile are used as negatives
"Unsupervised node clustering via contrastive hard sampling" (Cui et al., 2024)	Unsupervised node clustering	Virtual nodes interpolate class-dependent features while copying class-invariant features, producing competitive negatives near cluster boundaries
"Towards Robust Recommendation via Decision Boundary-aware Graph Contrastive Learning" (Tang et al., 2024)	Recommendation	Instance-wise perturbations are capped by the maximum admissible perturbation that does not cross the decision boundary
"BoundAD: Boundary-Aware Negative Generation for Time Series Anomaly Detection" (Wang et al., 18 Mar 2026)	Time-series anomaly detection	Reconstruction-model updates are controlled so that reconstructed windows shift toward the boundary of the normal manifold

In Contextrast, the boundary-aware component is called Boundary-Aware Negative (BANE) sampling rather than BACS (Sung et al., 2024). For each class $p_t$8 and layer $p_t$9, the method constructs a binary error map

$r_i$0

extracts the boundary set $r_i$1 of the mis-prediction region, computes the distance transform

$r_i$2

collects candidate negatives from mis-predicted pixels of class $r_i$3, sorts them by ascending $r_i$4, and selects the lowest-$r_i$5-percentile as the boundary-aware negative set. The resulting negatives are inserted into a multi-scale pixel-anchor loss at every encoder scale.

In MeCoLe, the paper does not define an explicit distance-to-boundary function, but it constructs hard negatives by synthesizing a virtual node $r_i$6 whose class-dependent embedding is interpolated between anchor node $r_i$7 and a node $r_i$8 from a different cluster,

$r_i$9

while copying the class-invariant part,

$\phi$0

Negatives are then drawn from predicted neighbors of $\phi$1 that are not already adjacent to $\phi$2 (Cui et al., 2024). The paper’s exposition states that $\phi$3 tends to place $\phi$4 in the mid-region between clusters.

In RGCL, boundary awareness is defined through the maximum admissible perturbation $\phi$5 that preserves the original preference ranking (Tang et al., 2024). Random perturbations are sampled and then projected into an instance-wise box determined by $\phi$6, ensuring that no augmentation exceeds the boundary-aware budget. Adversarial-contrastive views are then constructed by maximizing an InfoNCE-style loss with respect to additional noise under the same budget constraint.

In BoundAD, boundary-aware negative generation arises from reconstruction dynamics rather than explicit labeling or perturbation bounds (Wang et al., 18 Mar 2026). The model updates the reconstruction parameters according to a reinforcement-learning-controlled signed step

$\phi$7

and generates negatives as

$\phi$8

These are described as boundary-shifted versions of real normal windows.

Taken together, these methods indicate that “boundary-aware” can refer to at least four distinct constructions: geometric proximity to segmentation error contours, interpolation near cluster separators, perturbation budgets induced by recommendation decision boundaries, and trajectories near the normal data manifold. The shared structure is not a single equation but the use of hard negatives that are intentionally close to the model’s present decision frontier.

5. Empirical evidence and ablation patterns

The empirical profile of BACS in C-MOP is given through both full-system and component-wise evaluations (Yan et al., 11 Feb 2026). The paper reports that C-MOP consistently outperforms SOTA baselines like PromptWizard and ProTeGi, yielding average gains of $\phi$9 and $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$0. It further states that C-MOP enables a general LLM with 3B activated parameters to surpass a 70B domain-specific dense LLM.

The specific contribution of BACS is isolated in ablations. Replacing BACS with random failure sampling while keeping MGSC drops average accuracy from $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$1 to $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$2. Using naive error-only clustering yields $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$3. Applying BACS without MGSC yields $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$4. In the fine-grained sampling study, Hard Negatives alone achieve $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$5, Hard Negatives plus Anchors yield $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$6, Hard Negatives plus Boundary Pairs yield $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$7, and full BACS with all three feature types yields $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$8 (Yan et al., 11 Feb 2026). The paper interprets these results as showing that Anchors anchor correct logic, Hard Negatives capture systematic failures, and Boundary Pairs supply the highest-contrast signals.

Comparable empirical patterns appear in adjacent work. In Contextrast, an ablation on Cityscapes with an HRNet backbone shows that adding only CCL raised val-mIoU from $e_i = \phi(\mathrm{concat}(q_i, a_i, r_i)).$9 to $B$0, substituting semi-hard sampling bumped it to $B$1, and BANE sampling gave $B$2. On CamVid with HRNet, BANE sampling raised mIoU from $B$3 to $B$4, compared to $B$5 with plain CCL (Sung et al., 2024). Boundary-specific evaluation further shows stronger gains in B-mIoU than random multi-scale sampling.

In MeCoLe, the full method outperforms all 9 baselines on every dataset listed in the paper, with examples including Wikipedia $B$6 versus CommDGI $B$7, GRACE $B$8, and InfoGCL $B$9, and IMDB $K$0 versus the best baseline $K$1 (Cui et al., 2024). Ablations show that removing hard negative sampling causes large accuracy drops.

In RGCL, the paper reports superiority over all baselines on every dataset and metric, with ML-1M NDCG@10 improving from SimGCL $K$2 to RGCL $K$3, and larger gains on sparse settings such as Alibaba and Kuaishou (Tang et al., 2024). Removing boundary constraints leads to a $K$4 decrease in Recall@20 on the ablation benchmark cited in the summary.

In BoundAD, RL-guided boundary negatives improve ROC-AUC over random-noise and static-injection baselines, with Point AUC $K$5 and Seasonal AUC $K$6, and the seasonal false positive rate reported as approximately $K$7, compared with approximately $K$8 for random noise and approximately $K$9 for static injection (Wang et al., 18 Mar 2026).

Across these studies, the recurring empirical pattern is that boundary-aware negatives outperform random or weakly structured negatives, and that gains are often especially visible in settings where the task is dominated by local ambiguity: object edges, cluster separators, sparse recommendation neighborhoods, or the frontier of the normal manifold.

6. Interpretation, misconceptions, and methodological significance

A common misconception is that boundary-aware contrastive sampling necessarily requires an explicit distance-to-boundary function. The literature summarized here does not support that claim. Contextrast defines a literal distance transform over error boundaries (Sung et al., 2024), but MeCoLe explicitly does not define an explicit “distance to boundary” function, instead using virtual-node interpolation and predicted neighborhoods (Cui et al., 2024). RGCL defines boundary awareness through the largest perturbation that does not cross a ranking boundary (Tang et al., 2024), and BoundAD induces boundary-neighborhood samples through reconstruction trajectories regulated by reinforcement learning (Wang et al., 18 Mar 2026).

A second misconception is that BACS is a universally standardized module name. The evidence is narrower. BACS is explicitly named in C-MOP (Yan et al., 11 Feb 2026). By contrast, the node-clustering summary states that the corresponding paper does not actually introduce a module called “Boundary-Aware Contrastive Sampling” by that name (Cui et al., 2024). This suggests that the label should be used precisely: as a named component in C-MOP, and as a descriptive umbrella only when the usage is clearly marked as reinterpretation.

Methodologically, the significance of boundary-aware sampling lies in how it changes the gradient signal. Contextrast attributes its gains to hardness, fine detail, and multi-scale consistency around mis-segmented edges (Sung et al., 2024). RGCL notes that InfoNCE produces larger gradients for hard pairs, so pushing augmentations toward the decision boundary increases the usefulness of contrastive learning (Tang et al., 2024). BoundAD similarly emphasizes that negatives should preserve temporal semantic consistency while providing effective decision-boundary supervision (Wang et al., 18 Mar 2026). In C-MOP, the local, boundary-focused contrasts are said to reduce gradient variance and steer prompt evolution toward precise, impactful edits rather than noisy, broad-brush updates (Yan et al., 11 Feb 2026).

The broader implication is that boundary-aware contrastive sampling is less a single recipe than a design principle: negatives should be neither arbitrary nor excessively distant, but should remain semantically valid while approaching the regions in representation space where the model is most liable to confuse alternatives. Where that principle is implemented successfully, the reported effects are consistent: stronger discrimination, cleaner optimization signals, and improved robustness at the task’s most failure-prone interface.