Dynamic Experts Search (DES)

Updated 13 July 2026

Dynamic Experts Search (DES) is a conditional expert selection paradigm that adaptively chooses relevant models based on query context and local competence estimation.
It employs methodologies from classical ensemble selection to modern sparse MiE routing, balancing model accuracy with resource efficiency.
DES is applied in domains such as multi-domain adaptation, real-time object detection, and wireless-edge systems, demonstrating improved performance and reduced computation.

Dynamic Experts Search (DES) denotes a family of instance-adaptive selection methods in which the set of experts used for a decision is chosen online from the current query, token, domain, or runtime context rather than fixed globally. In classical multiple-classifier systems, the closely related term Dynamic Ensemble Selection refers to choosing one classifier or a locally competent subset for each query sample (Cruz et al., 2018, Cruz et al., 2018). In sparse Mixture-of-Experts (MoE) models, DES has also come to denote adaptive expert routing, dynamic expert-count control at test time, domain-specific expert specialization during fine-tuning, and resource-aware expert allocation under systems or communication constraints (Han et al., 26 Sep 2025, Huang et al., 2024, Li et al., 21 Sep 2025, Qin et al., 17 Mar 2025).

1. Scope and terminology

The term DES is used across several adjacent literatures, but the common structure is stable: a pool of experts exists, competence or utility is estimated conditionally, and selection is performed locally rather than by a single global rule. What changes across subfields is the identity of the “expert,” the evidence used for selection, and the objective being optimized.

Family	Searched object	Representative papers
Classical DES/DCS	Base classifier or local classifier subset	(Cruz et al., 2018, Cruz et al., 2018)
MoE routing and TTS	Neural experts or expert counts per token/path	(Huang et al., 2024, Lu et al., 23 Jul 2025, Han et al., 26 Sep 2025)
Domain/resource-aware DES	Domain-specialized experts or low-cost expert subsets	(Li et al., 21 Sep 2025, Qin et al., 17 Mar 2025)
Broader expert retrieval/composition	Expert-labeled cases or expert checkpoints	(Zlabinger et al., 2020, Feng et al., 2024)

In the classical formulation, DES is local model selection in a multiple-classifier system. In modern MoE usage, DES often means conditional subnet selection inside a neural network. A broader interpretation also appears in systems that retrieve expert-produced examples or search over expert checkpoints. This suggests that DES is best understood as a general conditional-expert-selection paradigm rather than a single algorithmic family.

2. Classical dynamic ensemble selection

The software and methodological baseline for classical DES is codified by DESlib, which separates Dynamic Classifier Selection (DCS) from Dynamic Ensemble Selection (DES) and organizes the workflow around three components: definition of the region of competence, the information source for competence estimation, and the selection rule (Cruz et al., 2018). Operationally, the pipeline is: train a pool of base classifiers, define the query’s local neighborhood in a dynamic selection dataset, estimate each classifier’s competence in that region, select either one classifier (DCS) or a subset (DES), and aggregate predictions. DESlib exposes this structure through dcs, des, and static modules and implements methods such as META-DES, KNORA-E, KNORA-U, DES-P, KNOP, OLA, LCA, and static baselines including Single Best, Static Selection, Oracle, and Stacked Generalization (Cruz et al., 2018).

META-DES formalized classical DES as a meta-learning problem rather than a single local-accuracy heuristic. For each classifier–query pair, it constructs a meta-feature vector from five criteria: neighbors’ hard classification behavior, posterior probabilities in the region of competence, overall local accuracy, output-profile behavior in decision space, and classifier confidence relative to the decision boundary (Cruz et al., 2018). The framework uses a region of competence of size $K=7$ , $K_p=5$ similar output profiles, a consensus threshold $h_C=70\%$ , and trains a meta-classifier to decide whether a classifier is competent enough to join the local ensemble (Cruz et al., 2018). The importance of META-DES is conceptual as much as empirical: competence estimation becomes a learned local inference problem rather than a fixed statistic.

A major line of DES research concerns the quality of the local evidence used to estimate competence. PS-DES adds a second decision layer: instead of trusting one DES technique globally, it runs several DES methods for the same query and then post-selects the candidate ensemble with highest estimated potential (Cordeiro et al., 2023). In the reported implementation, the candidate set consists of KNORA-U, KNOP, DES-P, and META-DES, all using the same region of competence; among the proposed post-selection metrics, PS-DES-acc obtains the best average rankings for accuracy, F-score, and MCC across 20 UCI datasets (Cordeiro et al., 2023). The central idea is that dynamic search can itself be hierarchical: first over base classifiers, then over DES strategies.

FIRE-DES++ addresses a different failure mode: a poor region of competence can make even a reasonable competence estimator unreliable. It augments frienemy-based online pruning with two changes: ENN filtering to denoise the dynamic selection set and KNNE to enforce class-representative local neighborhoods (Cruz et al., 2018). In the eight-scenario ablation reported in the paper, the full configuration combining ENN, KNNE, and DFP achieves the best average rank (3.75) and mean AUC (82.95), whereas the original FIRE-DES configuration is the worst-ranked scenario among the eight (Cruz et al., 2018). Across eight KNN-based DES techniques, FIRE-DES++ improves seven relative to FIRE-DES and also outperforms stronger DES baselines such as META-DES, META-DES.Oracle, and RRC in that experimental setting (Cruz et al., 2018).

FH-DES replaces KNN neighborhoods altogether with fuzzy hyperboxes that model competence or incompetence regions directly (Davtalab et al., 2022). The two variants are FH-DES-C, built from correctly classified samples, and FH-DES-M, built from misclassified samples; the latter is emphasized as novel because it estimates competence indirectly by modeling where a classifier fails (Davtalab et al., 2022). In the comparative study, FH-DES-M achieves the best average accuracy (81.89) among the reported FH-DES variants and the highest average accuracy among the compared DES methods, while using many fewer hyperboxes than raw DSEL instances in large-scale settings (Davtalab et al., 2022). The same scalability concern motivates IncA-DES for nonstationary streams, which couples KNORAE with a drift-aware DSEW window, an overlap-based filter that skips DES when local consensus is high, and an Online K-d tree for faster neighborhood search under concept drift (Barboza et al., 16 Jul 2025). With RDDM, pool size $D=75$ , $F=200$ instances per classifier, $k=5$ , and overlap threshold $\omega=0.8$ , IncA-DES attains the best average accuracy against seven state-of-the-art stream baselines and the smallest processing time among the most accurate methods (Barboza et al., 16 Jul 2025).

4. MoE routing as modern DES

In sparse MoE models, DES becomes token-wise or path-wise expert routing. One direct formulation is confidence-thresholded dynamic routing: instead of fixed Top- $K$ , the model activates the smallest prefix of ranked experts whose cumulative routing probability exceeds a threshold $p$ (Huang et al., 2024). In "Harder Tasks Need More Experts," the adaptive rule is

$t = \min \left\{ k \in \{1,\dots,N\} \;\middle|\; \sum_{j=1}^{k} P_{I_j} \ge p \right\},$

so harder or more ambiguous inputs, identified by flatter router distributions, receive more experts (Huang et al., 2024). The model combines standard language-model loss with a load-balance term and an entropy-based dynamic loss, and reports an average gain of 0.7 points over fixed Top-2 while using less than 90% activated parameters relative to Top-2 (Huang et al., 2024). The layerwise analysis is also DES-relevant: lower layers use more experts, upper layers trend toward one expert, and BBH triggers the largest average expert count (1.87) among the downstream benchmarks (Huang et al., 2024).

Dynamic-DINO extends this logic to real-time open-vocabulary object detection by decomposing decoder FFNs into fine-grained experts and routing tokens through sparse MoE blocks (Lu et al., 23 Jul 2025). Its search space is explicitly

$K_p=5$ 0

where $K_p=5$ 1 is the number of experts, $K_p=5$ 2 the number activated per layer, and $K_p=5$ 3 the number of layers (Lu et al., 23 Jul 2025). The method preserves the original dense FFN exactly at initialization through weight allocation and router replication, then learns token-wise hard top- $K_p=5$ 4 routing over the expanded expert set (Lu et al., 23 Jul 2025). The reported routing analysis shows shallow layers cooperating with diverse peers and deeper layers forming stable collaborations in which each expert keeps roughly 2–3 fixed partners (Lu et al., 23 Jul 2025). At the systems level, active parameters remain constant at 178M while total parameters rise to 184M, 197M, and 222M for larger Dynamic-DINO variants (Lu et al., 23 Jul 2025).

The most literal use of the phrase appears in "Dynamic Experts Search: Enhancing Reasoning in Mixture-of-Experts LLMs at Test Time," which treats the number of activated experts as a search variable inside verifier-guided test-time scaling (Han et al., 26 Sep 2025). DES introduces Dynamic MoE, which conditions generation on a chosen expert count $K_p=5$ 5, and Expert Configuration Inheritance, which keeps $K_p=5$ 6 fixed along a reasoning trajectory while allowing different trajectories to use different counts (Han et al., 26 Sep 2025). The search state is a pair $K_p=5$ 7, initialized over expert counts $K_p=5$ 8, and top- $K_p=5$ 9 candidates are retained by verifier score at each reasoning step (Han et al., 26 Sep 2025). On Qwen3-30B-A3B-Instruct with Qwen2.5-Math-PRM-7B and $h_C=70\%$ 0, DES reports 93.20 accuracy on MATH500 and 94.51 on HumanEval, exceeding Best-of-N and usually BeamSearch or DVTS at comparable token cost (Han et al., 26 Sep 2025).

5. Domain-aware and resource-aware DES

DES can also operate at the domain or systems-allocation level rather than only per token. DES-MoE uses Dynamic Expert Specialization for multi-domain adaptation of a pretrained MoE model, combining an adaptive lightweight router regularized by distillation, online domain–expert affinity tracking, and a three-phase schedule of Warm-Up, Stabilization, and Consolidation (Li et al., 21 Sep 2025). The key specialization statistic is

$h_C=70\%$ 1

which is thresholded with $h_C=70\%$ 2 and updated online with momentum $h_C=70\%$ 3 to build domain-specific expert masks (Li et al., 21 Sep 2025). Gradients are then restricted to domain-relevant experts, and experts active for multiple domains can be duplicated to isolate specialization (Li et al., 21 Sep 2025). In the six-domain study, DES-MoE reaches 51.6 average downstream score, compared with 46.6 for mixed FFT and 47.7 for mixed ESFT-Gate, while preserving 62.4 average general ability, equal to the vanilla LM baseline reported in the paper (Li et al., 21 Sep 2025). The abstract further reports 89% forgetting reduction relative to full fine-tuning as domains scale from 2 to 6 and 68% faster convergence (Li et al., 21 Sep 2025).

A different resource-aware formulation appears in distributed wireless-edge MoE. There, expert choice must satisfy both task relevance and channel conditions, so the DES objective minimizes communication and computation cost subject to a gating-score QoS constraint (Qin et al., 17 Mar 2025). The resulting problem is NP-hard, and the proposed DES algorithm uses branch-and-bound with a linear-relaxed lower bound driven by the energy-to-score ratio $h_C=70\%$ 4 (Qin et al., 17 Mar 2025). The full framework is extended to JESA, which jointly optimizes expert selection and OFDMA subcarrier allocation; under the paper’s i.i.d. subcarrier-rate assumption, JESA is asymptotically optimal as the number of subcarriers grows (Qin et al., 17 Mar 2025). Numerically, the framework can reduce cost by up to 50% relative to Top- $h_C=70\%$ 5, and the paper highlights a setting with $h_C=70\%$ 6 that yields about 40% energy reduction for 5% accuracy loss relative to Top-2 (Qin et al., 17 Mar 2025). In this formulation, DES is not merely semantic routing but a constrained optimization over expert utility and communication quality.

6. Systems substrate, broader interpretations, and recurring limitations

A recurrent theme in modern DES is that selecting experts well is not sufficient if executing the resulting dynamic pattern is inefficient. DynaMoE is explicitly a systems paper rather than a new routing algorithm: it keeps standard MoE routing and losses, but introduces dynamic recompilation so the computation graph can adapt to runtime expert loads without a cold restart (Kossmann et al., 2022). Its MoE-specific operators include TopK, GroupBy, Aggregate, and AggregateSpec, and its runtime exploits a launch frontier/execution frontier separation to overlap trigger computation and remapping with ongoing GPU work (Kossmann et al., 2022). The two main DES-relevant optimizations are dynamic capacity factors, which resize expert capacities according to realized routing loads, and sample assignment caching, which exploits stabilized routing to decouple gate and expert execution (Kossmann et al., 2022). Relative to FastMoE, DynaMoE reports up to 1.8x speedup and support for 2.3x larger models before out-of-memory; dynamic capacities alone give 1.4x speedup, and adding caching raises total speedup to 1.7x while preserving accuracy (Kossmann et al., 2022). This clarifies an important distinction: DES may refer either to deciding which experts should run or to making those dynamic decisions executable at low overhead.

The concept also extends beyond classifier pools and neural routers. DEXA performs dynamic retrieval over a repository of expert-labeled biomedical annotation examples: for each target sentence, it retrieves the top 3 semantically similar expert-labeled sentences using BioSent2Vec and presents them to crowd workers (Zlabinger et al., 2020). In the reported PIO study, this raises average worker agreement with experts to 0.68 from 0.40, and three-vote majority aggregation reaches 0.780/0.757/0.694 for P/I/O (Zlabinger et al., 2020). Model Swarms generalizes the notion further by searching over LLM expert checkpoints in weight space with a swarm-intelligence update rule guided by a utility function (Feng et al., 2024). Using a pool of expert checkpoints or LoRA adapters, it improves over 12 composition baselines by up to 21.0% across tasks and contexts and is reported to work with as few as 200 examples in low-data adaptation settings (Feng et al., 2024). These cases suggest that DES can refer to online retrieval of expert-produced evidence or dynamic search over expert parameterizations, not only per-token gating.

Across these literatures, several limitations recur. In classical DES, competence estimation remains sensitive to the quality of the local evidence, whether represented by KNN regions, frienemy pairs, or hyperboxes, and inference can become expensive without careful indexing or filtering (Cruz et al., 2018, Davtalab et al., 2022, Barboza et al., 16 Jul 2025). In MoE DES, threshold policies, routing stability, expert-count schedules, and verifier quality are often heuristic, and hardware efficiency can lag behind algorithmic sparsity when expert dispatch is sequential or poorly mapped (Huang et al., 2024, Lu et al., 23 Jul 2025, Han et al., 26 Sep 2025). In multi-domain and resource-aware settings, the method may depend on explicit domain labels, accurate channel models, or stable router statistics (Li et al., 21 Sep 2025, Qin et al., 17 Mar 2025). This suggests that the central problem of DES is no longer only identifying the semantically appropriate expert, but jointly managing semantic relevance, optimization stability, and execution cost.