Diversity-driven Distribution Calibration (DDC)

• DDC is a family of methods that use diversity to adjust a learned distribution toward a target structure for improved representativeness and calibration.
• In active finetuning (ActiveDC), diversity is introduced via prototype repulsion and pseudo-label augmentation to reduce distribution bias, yielding significant accuracy gains.
• Across applications in language models, object detection, and vision-language tasks, DDC strategies reallocate probability mass and optimize geometric dispersion to better capture valid outputs.

Searching arXiv for the papers on arXiv and closely related work on diversity-driven distribution calibration. arxiv_search(query="(Xu et al., 2023)", max_results=5) arxiv_search(query="(Xu et al., 2023) ActiveDC Distribution Calibration for Active Finetuning", max_results=5) Diversity-driven Distribution Calibration (DDC) denotes a family of methods and analytical perspectives in which diversity is used to correct a mismatch between a learned or selected distribution and a target structure. In the cited literature, that target structure varies by domain: the full unlabeled pool in active finetuning, the valid set of continuations in autoregressive decoding, the confidence–IoU relation in object detection, or the angular arrangement of class text features in vision-LLMs. The term is not used as a single standardized method name across these works; rather, it functions as an umbrella description for approaches that treat diversity as a mechanism for improving representativeness, validity-aware coverage, uncertainty quality, or calibration (Xu et al., 2023, Banayeeanzade et al., 11 May 2026, Syed et al., 2024, Ahamed et al., 30 Oct 2025, Pace et al., 2020).

1. Conceptual scope

A common structure recurs across the literature. First, some distribution is judged to be deficient: an actively selected labeled subset is too narrow, token probabilities collapse onto a few continuations, detector branches become too similar, or prompt tuning clusters class-text features into poorly dispersed directions. Second, a diversity-inducing mechanism is introduced. Third, that mechanism is used not as an end in itself, but to improve calibration, coverage, or robustness under a downstream constraint such as annotation budget, output validity, OOD sensitivity, or confidence reliability.

Paper	Distributional object	Diversity mechanism
"ActiveDC: Distribution Calibration for Active Finetuning" (Xu et al., 2023)	Selected subset vs. full unlabeled pool	Prototype repulsion plus category-aware calibration
"Sampling More, Getting Less: Calibration is the Diversity Bottleneck in LLMs" (Banayeeanzade et al., 11 May 2026)	Token distribution over valid and invalid continuations	Reallocation of mass to fix order and shape calibration
"Situation Monitor: Diversity-Driven Zero-Shot Out-of-Distribution Detection using Budding Ensemble Architecture for Object Detection" (Syed et al., 2024)	Confidence and uncertainty behavior in detection	Cosine-similarity diversity between tandem heads
"A-TPT: Angular Diversity Calibration Properties for Test-Time Prompt Tuning of Vision-LLMs" (Ahamed et al., 30 Oct 2025)	Distribution of normalized class text features	Maximization of minimum pairwise angular distance
"Learning Diverse Representations for Fast Adaptation to Distribution Shift" (Pace et al., 2020)	Label-conditional representation dependence across models	Minimization of conditional total correlation

The literature also makes clear that DDC is not reducible to a single calibration primitive. In ActiveDC, calibration is a pre-finetuning data augmentation and relabeling step, not a finetuning loss. In the LLM setting, the emphasis is diagnostic and theoretical rather than algorithmic. In Situation Monitor, calibration is induced during training rather than by post-hoc scaling. In A-TPT, what is regularized is the geometry of textual prototypes on the unit hypersphere rather than the output distribution directly (Xu et al., 2023, Banayeeanzade et al., 11 May 2026, Syed et al., 2024, Ahamed et al., 30 Oct 2025).

2. Active finetuning as subset-level distribution calibration

The clearest direct instantiation of DDC appears in ActiveDC for active finetuning. The setting assumes a pretrained encoder

$$\mathcal{M}(\cdot;\omega_0): \mathcal{X}\rightarrow \mathbb{F}^u,$$

an unlabeled pool

$\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$

and a fixed annotation budget $B$. The problem is to select $B$ samples for oracle labeling and then finetune using that limited labeled set. The paper’s motivation is that under tight budgets the selected subset is a poor empirical approximation to the full pool distribution, inducing distribution bias: some modes, classes, or feature-space regions are underrepresented or missing, and the finetuned model overfits to the narrow selected subset (Xu et al., 2023).

ActiveDC decomposes the remedy into two stages: data selection and distribution calibration. In the first stage it inherits the ActiveFT selection engine and optimizes distribution discrepancy plus diversity in continuous feature space. With L2-normalized features and cosine similarity, the continuous prototype objective is

$$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$

and the optimization loss is

$$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$

The first term pulls prototypes toward the full pool distribution, while the second discourages collapse and promotes coverage. In the paper’s own DDC-oriented synthesis, diversity is explicit here through the repulsion term and implicit later through category recovery from unlabeled structure (Xu et al., 2023).

The second stage is the distinctive calibration mechanism. After oracle labeling of the selected subset and before supervised finetuning, ActiveDC augments the small labeled set with pseudo-labeled samples chosen from the unlabeled pool. The method applies Tukey’s Ladder of Powers transformation,

$$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$

clusters transformed unlabeled features into $K$ groups using K-Means, aligns cluster identities to semantic categories via the labeled features, computes pseudo-category statistics, and then calibrates those statistics using the small true labeled set. For class $i$, the calibrated mean and covariance are

$$\mu _i = (1-\beta) \hat{\mu}_i + \beta \mu_i^l,\qquad \sigma^2_i = \mathrm{S} ^2_i + \xi,$$

with

$\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$0

Pseudo-features are sampled from $\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$1, then replaced by nearest real unlabeled features, which are assigned pseudo-labels and added to the labeled pool. A filtering step then uses Earth Mover’s Distance between the extended subset distribution and the full pool distribution to remove harmful pseudo-labeled additions (Xu et al., 2023).

This procedure is DDC in a strong sense because it calibrates a scarce labeled subset toward the global unlabeled distribution while preserving category-wise coverage. The paper explicitly states that a generic DDC idea would be to select or synthesize a labeled subset whose empirical distribution matches the global unlabeled distribution while ensuring diverse coverage of categories and modes, and it interprets ActiveDC as a two-stage diversity-aware distribution calibration method. Empirically, the gains are largest in low-budget regimes: for example, against ActiveFT, Top-1 accuracy improves on CIFAR100 from $\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$2 at $\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$3 and from $\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$4 at $\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$5, and on ImageNet from $\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$6 at $\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$7 (Xu et al., 2023).

3. Validity-aware token calibration in LLMs

In autoregressive language modeling, DDC appears as a theoretical diagnosis of diversity collapse. The central claim is that diversity collapse arises from how probability mass is allocated during decoding, specifically through two forms of miscalibration. Order calibration fails when valid next tokens are not reliably ranked above invalid ones. Shape calibration fails when mass is over-concentrated on a few valid continuations while a heavy tail mixes low-probability valid and invalid tokens. The paper formalizes generation with

$\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$8

defines the valid set $\mathcal{P}^u=\{x_i\}_{i\in[N]}\sim p_u,$9, the number of valid continuations

$B$0

and the valid next-token set

$B$1

and then defines validity and diversity over the induced sequence distribution (Banayeeanzade et al., 11 May 2026).

The formal DDC-relevant targets are particularly sharp. Order calibration is defined by

$B$2

for any valid token $B$3 and invalid token $B$4. Shape calibration is defined by

$B$5

The first criterion concerns rank-separation of valid and invalid continuations; the second concerns the allocation of mass within the valid region in proportion to downstream valid support. This matters because the paper argues that top-$B$6, top-$B$7, min-$B$8, and related rank-based rules cannot recover diversity if valid and invalid tokens are interleaved in rank, while temperature scaling cannot selectively flatten the valid head without inflating the heavy mixed tail (Banayeeanzade et al., 11 May 2026).

The local-to-global analysis gives a formal reason that this is a calibration problem rather than merely a decoding-heuristic problem. The appendix proves exact multiplicative decompositions of sequence precision and sequence recall, implying that local calibration failures compound over generation length. Under a stylized shape-miscalibration model, the paper further proves that if validity is kept high, diversity decays exponentially in branching length. The DDC implication drawn in the synthesis is that a successful method should separate or estimate valid versus invalid continuations and then redistribute mass within the valid region rather than flattening indiscriminately (Banayeeanzade et al., 11 May 2026).

A critical limitation is explicit: the paper does not propose a direct calibration algorithm that fully resolves the bottleneck. It offers controlled diagnostics, oracle cutoff baselines, and a theoretical foundation for calibration-aware decoding, including tasks with exactly known valid sets such as unconstrained random number generation, constrained random number generation, and random US state generation. Across 14 LLMs spanning multiple families and scales, it concludes that diversity collapse is not merely a limitation of particular sampling heuristics, but a consequence of order and shape miscalibration in the LLM distribution (Banayeeanzade et al., 11 May 2026).

4. Training-time confidence calibration in object detection

In transformer-based object detection, DDC takes the form of structured ensemble diversity that simultaneously supports zero-shot OOD detection and confidence calibration. Situation Monitor is built on DINO-DETR and modifies the Budding Ensemble Architecture by keeping the shared backbone, transformer encoder, and transformer decoder, but duplicating the final 3-layer feed-forward regression heads after the decoder to create two tandem detectors, $B$9 and $B$0. The paper’s main novelty is to add a diversity regularizer so that the tandem branches do not collapse to nearly identical solutions, which would weaken both uncertainty estimation and OOD detection (Syed et al., 2024).

The DBEA training objective is

$B$1

where $B$2. The tandem-aiding component reduces disagreement on positive detections, the tandem-quelling component increases disagreement on no-object predictions, and the diversity term minimizes cosine similarity between branch classification logits: $B$3 Because this similarity is minimized, the effect is to encourage lower similarity and hence higher diversity between branch logits. The paper explicitly describes this as a learned calibration mechanism during training rather than a post-hoc scaling procedure (Syed et al., 2024).

Calibration is evaluated not by ECE but by Pearson correlation between confidence scores and IoU, using PCorr(all) and PCorr(tp). On KITTI-trained DINO-DETR, the reported values are PCorr(all) $B$4 and PCorr(tp) $B$5 for Vanilla-DINO-DETR, $B$6 and $B$7 for BEA-DINO-DETR, and $B$8 and $B$9 for DBEA-DINO-DETR. On BDD100K-trained DINO-DETR, Vanilla-DINO-DETR has PCorr(all) $$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$0 and PCorr(tp) $$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$1, while DBEA-DINO-DETR reaches $$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$2 and $$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$3. The method also substantially improves Far-OOD detection, with KITTI-trained DBEA-DINO-DETR reaching AUROC $$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$4 and BDD100K-trained DBEA-DINO-DETR reaching AUROC $$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$5 against COCO as Far-OOD (Syed et al., 2024).

This is DDC in the sense that diversity is deliberately injected to shape both predictive confidence and uncertainty behavior. The paper is careful, however, that Situation Monitor is primarily an OOD detection method rather than a generic calibration method. It also evaluates calibration only through Pearson correlation with IoU, not through ECE, localization-aware ECE, risk-coverage curves, or reliability diagrams. A plausible implication is that this work is best read as a diversity-driven, training-integrated calibration mechanism for confidence and localization quality, but not as a full predictive-distribution calibration framework (Syed et al., 2024).

5. Angular diversity and latent prototype calibration in vision-LLMs

In test-time prompt tuning for vision-LLMs, DDC appears as geometric calibration of class-text features. The setting is CLIP-style zero-shot classification, where a prompt for class $$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$6 is encoded as $$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$7, an image $$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$8 is encoded as $$\theta_{\mathcal{S},opt} = \underset{\theta_\mathcal{S}}{\arg\min}\; D(p_{f_u},p_{\theta_\mathcal{S}}) - \lambda R(\theta_\mathcal{S}) \quad s.t.\quad \|\theta^j_\mathcal{S}\|_2=1,$$9, and classification uses cosine similarity followed by a softmax with fixed temperature $$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$0. The paper argues that standard TPT often improves accuracy but worsens calibration because the adapted prompt embeddings reduce dispersion between textual features, producing overconfident predictions (Ahamed et al., 30 Oct 2025).

A-TPT addresses this by enforcing angular diversity among normalized class-text features on the unit hypersphere. Let $$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$1 be the class-text feature matrix and $$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$2 its row-wise normalization. The cosine-similarity matrix is

$$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$3

the angular distance matrix is

$$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$4

and angular diversity is defined as

$$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$5

The regularizer is $$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$6, and the full adaptation objective is

$$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$7

This objective maximizes the smallest pairwise angular distance for each class and thereby approximates a best-packing configuration on the hypersphere (Ahamed et al., 30 Oct 2025).

The paper’s theoretical arguments explain why this is calibration-relevant. It invokes the Tammes problem as geometric motivation, argues that uniformity on the hypersphere is more appropriate than average Euclidean dispersion for normalized CLIP features, and shows that an orthogonality-based alternative suffers from vanishing gradients when features are nearly colinear. By contrast, the gradient norm for the angular distance objective is independent of the angle, which the paper presents as a practical reason for better optimization in the crowded regime (Ahamed et al., 30 Oct 2025).

Empirically, the method reduces calibration error while preserving comparable accuracy. For CLIP ViT-B/16 on the main fine-grained benchmark average, ECE changes from $$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$8 for the baseline, $$\begin{split} L &= - \underset{f_i \in \mathcal{F}^u}{E} \left [ \frac{cos(f_i,\theta^{c_i}_\mathcal{S})}{\tau} \right ] + \underset{j \in [B]}{E} \left [ \log \sum_{k \ne j,k \in [B]} \exp\!\left(\frac{cos(\theta^{j}_\mathcal{S},\theta^{k}_\mathcal{S})}{\tau}\right) \right]. \end{split}$$9 for TPT, $$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$0 for C-TPT, and $$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$1 for O-TPT to $$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$2 for A-TPT, with accuracy $$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$3. For CLIP RN50, the average ECE becomes $$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$4 for A-TPT. A-TPT also improves SCE, shows robustness across prompt initializations, and generalizes to medical datasets such as ISIC 2018, KatherColon, and Covid tasks. The paper therefore frames its contribution as prompt calibration through angular diversity, and under a DDC interpretation this is a latent-feature distribution calibration method rather than a direct output-probability recalibration method (Ahamed et al., 30 Oct 2025).

6. Relation to representation diversity, assumptions, and limitations

A broader precursor to DDC-like thinking appears in work on learning diverse representations for fast adaptation to distribution shift. That paper is explicit that it is not about calibration in the probabilistic sense. Instead, it trains multiple models $$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$5 and minimizes

$$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$6

where conditional total correlation is

$$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$7

The objective encourages label-conditionally distinct representations so that small labeled samples from the shifted distribution can identify or recombine mechanisms that remain valid after shift. This is directly relevant to the diversity side of DDC, but only indirectly relevant to calibration (Pace et al., 2020).

Across the literature, several common misconceptions are corrected. DDC is not simply “more sampling”: in the LLM analysis, the claim is that diversity collapse is caused by order and shape miscalibration in the model distribution, not merely by the choice of top-$$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$8, top-$$\hat{x} = \begin{cases} x^\lambda, & \text{if }\lambda \ne 0 \ \log x, & \text{if }\lambda = 0, \end{cases}$$9, min-$K$0, or temperature. DDC is not necessarily post-hoc calibration: Situation Monitor and A-TPT modify training or test-time adaptation objectives, while ActiveDC performs pre-finetuning data augmentation and relabeling. DDC is also not identical to generic ensemble variance: Situation Monitor requires diversity to be explicitly encouraged because branch collapse would otherwise make disagreement uninformative (Banayeeanzade et al., 11 May 2026, Syed et al., 2024, Xu et al., 2023, Ahamed et al., 30 Oct 2025).

The limitations are likewise domain-specific and explicit. ActiveDC depends on pretrained feature quality, the known downstream class count $K$1, clustering that roughly aligns with semantic categories, and pseudo-label retrieval that does not introduce too much noise; too many generated pseudo-samples degrade performance (Xu et al., 2023). The LLM validity–diversity framework is a diagnostic and theoretical foundation, not a practical decoder that fully resolves the problem (Banayeeanzade et al., 11 May 2026). Situation Monitor demonstrates calibration improvements only through Pearson correlation with IoU and is instantiated only on DINO-DETR for object detection (Syed et al., 2024). A-TPT’s natural-shift calibration improvements are stronger for ViT-B/16 than for RN50 in the reported averages, and the method operates in the specific geometry of normalized CLIP-style embeddings (Ahamed et al., 30 Oct 2025). The representation-diversity approach requires labeled data from the shifted distribution for adaptation and validation, and its experiments are on controlled Colored-MNIST-style benchmarks rather than natural large-scale shifts (Pace et al., 2020).

Taken together, these works suggest that DDC is best understood as a cross-domain design principle: identify a distributional bottleneck, introduce a diversity mechanism matched to that bottleneck, and use the resulting redistribution to improve representativeness, validity-aware coverage, uncertainty quality, or calibration. What changes across domains is the calibrated object—subset distribution, token probabilities, branch disagreements, text-feature geometry, or label-conditional representations—while the unifying idea remains that diversity is useful when it is structured toward a calibration target rather than maximized indiscriminately (Xu et al., 2023, Banayeeanzade et al., 11 May 2026, Syed et al., 2024, Ahamed et al., 30 Oct 2025, Pace et al., 2020).