Dynamic Experts Search: Enhancing Reasoning in Mixture-of-Experts LLMs at Test Time

1 INTRODUCTION

• Existing TTS methods rely on output-level sampling, while DES treats MoE expert activation as another search dimension.The introduction identifies this as an underexplored opportunity in MoE models.Source: §1 · pp. 1–2
• Varying activated expert counts leaves overall accuracy little changed while yielding complementary sets of correctly solved problems.The paper describes low Jaccard similarity between the solved-problem sets for different configurations.Source: §1 · pp. 1–2, Figure 1 · p. 2
• DES controls expert counts during inference and keeps each count consistent along a reasoning trajectory while varying counts across runs.These two design choices are Dynamic MoE and Expert Configuration Inheritance.Source: §1 · p. 2
• DES consistently outperforms existing TTS strategies at comparable cost across diverse MoE models, verifiers, and reasoning benchmarks.The reported gains span math, code, and knowledge benchmarks, with higher accuracy and precision.Source: §1 · p. 2

2 RELATED WORK

• Prior TTS approaches generate candidate responses and select answers through voting or optimization, with search, iterative refinement, and process-level rewards among the explored strategies.The section cites majority voting, search methods, and process-level rewards (Kang et al., 2024; Snell et al., 2024).Source: §2 · pp. 2–3
• The section identifies temperature sampling as a common source of output diversity and contrasts it with DES's broader exploration through MoE architectures.It also notes that Liu et al. (2025) found the optimal scaling strategy varies with policy models, PRMs, and problem difficulty.Source: §2 · p. 3
• MoE research has used sparse top-k routing to scale models and explored dynamic expert selection, while DES brings MoE flexibility into test-time search.The section cites GShard and Switch Transformer for sparse routing, and Huang et al. (2024), Li et al. (2023), Zeng et al. (2024), and Jin et al. (2025) for dynamic selection.Source: §2 · p. 3

3.1 PROBLEM DEFINITION

• In standard TTS search, a policy model generates candidate intermediate steps and a verifier selects the highest-reward action to extend the current state.The process stops at the maximum reasoning depth or when an end-of-sequence token appears.Source: §3.1 · p. 3
• Small changes in MoE expert counts have little effect on overall accuracy but can solve distinct, partially non-overlapping problem sets.This motivates using activation-count variation as a source of reasoning diversity.Source: §3.1 · p. 3

3.2 DYNAMIC EXPERTS SEARCH

• Unlike approaches that use a fixed expert count throughout search, DES explores varying counts through Dynamic MoE and Expert Configuration Inheritance.Together, the components form a strategy for dynamic expert selection during multi-step reasoning.Source: §3.2 · p. 4
• Dynamic MoE treats the number of activated experts as a tunable inference-time parameter and conditions each generated reasoning step on that count.For a state s and count k, the policy samples the next step from πθ(· | s, k).Source: §3.2 · p. 4
• Expert Configuration Inheritance keeps the expert count consistent along each reasoning trajectory, while verifier scores steer later search toward more promising counts.The search initially explores a predefined set of counts; maintaining configuration consistency supports implicit configuration-level credit assignment.Source: §3.2 · pp. 4–5
• DES pairs initial candidate steps with their expert counts, retains high-scoring states, propagates each count to the next step, and returns a highest-scoring or majority-vote answer.The search continues until the maximum step count or an end-of-sequence token; the computation budget is N rollouts per problem.Source: §3.2 · pp. 4–5

4.1 SETUP

• The setup evaluates four MoE policy models with two process reward model verifiers across math, code, and knowledge benchmarks.The reported metrics are accuracy, precision, and generated tokens per problem; baselines include Best-of-N, Beam Search, and Diverse Verifier Tree Search.Source: §4.1 · pp. 5–7

4.2 MAIN RESULT

• DES outperforms TTS baselines in accuracy and precision with computational cost comparable to the baselines.The experiments report accuracy and precision comparisons and measure cost by generated tokens per problem.Source: Table 1 · p. 6, §4.2 · p. 7, §4.3 · p. 8
• On Qwen3-30B-A3B, DES achieves comparable or better accuracy than thinking mode.The evaluated modes also include instruction tuning and DES combined with instruction tuning.Source: Table 2 · p. 7, §4.2 · p. 8
• DES outperforms baselines across nearly all tested budgets, and its performance gap widens as the budget N increases.The comparison uses OLMoE-1B-7B-Instruct with Llama3.1-8B-PRM-Deepseek-Data on GSM8K and MATH500.Source: Figure 3 · p. 6, §4.2 · p. 8

4.3 ABLATION STUDY

• DES does not activate more experts than the policy model's default configuration in the reported Qwen3 experiments.Figure 4 reports both overall average activated experts and averages at each timestep.Source: Figure 4 · p. 7, §4.3 · p. 8
• With an initial range of 4–11 experts, the average activated-expert count aligns with the initial values' mean of 7.5.The results suggest candidate expert counts are explored uniformly rather than the model simply favoring larger counts.Source: §4.3 · p. 8, Table 3 · p. 8
• Exploring different activated-expert counts with Dynamic MoE yields stable performance improvements over vanilla Beam Search across computational budgets.The ablation varies expert counts uniformly when generating candidate next reasoning steps on MATH500.Source: §4.3 · p. 9, Table 5 · p. 9
• Expert Configuration Inheritance improves both accuracy and pass@N compared with exploring different expert counts independently at each timestep.Pass@N measures the probability of obtaining a correct answer within N attempts under a fixed computational budget.Source: §4.3 · p. 9, Figure 5 · p. 9

5 LIMITATION

• DES requires an external verifier during search, introducing additional communication overhead.This is a limitation shared with other test-time scaling strategies.Source: §5 · p. 9
• DES performance partly depends on verifier quality and reliability.Misaligned evaluations can hinder final performance.Source: §5 · p. 9

6 CONCLUSION

• DES makes the number of activated experts a controllable search dimension in MoE models, enabling capacity-aware exploration.The conclusion describes DES as harnessing MoE modularity to explore complementary reasoning capabilities.Source: §6 · pp. 9–10
• Experiments across multiple reasoning benchmarks show DES consistently outperforming existing TTS baselines.Source: §6 · p. 10

SUMMARY OF THE APPENDIX

• The appendix covers additional experiments, broader impact, implementation, licenses, and disclosure of LLM use.Source: SUMMARY OF THE APPENDIX · p. 14

A.1 EXPERIMENTS USING OTHER MODELS

• DES substantially outperforms other search strategies on SVAMP and GSM8K when evaluated with smaller models.The evaluation uses Qwen2.5-Math-PRM-7B as verifier; the authors selected these benchmarks because the models struggled with AIME AI-MO (2024; 2025).Source: §A.1 · p. 14
• The evaluation uses DeepSeek-V2-Lite-Chat and OLMoE-1B-7B-Instruct, with Qwen2.5-Math-PRM-7B as verifier.Source: §A.1 · p. 14
• Table 6 reports accuracy and precision across strategies for the Qwen2.5-Math-PRM-7B policy model, using 64 rollouts per problem.Source: Table 6 · p. 14

A.2 ADDITIONAL ABLATION STUDY

• With Llama3.1-8B-PRM-Deepseek-Data as verifier, DES still outperforms the baseline search method.Table 7 reports accuracy and precision across strategies using 64 rollouts per problem.Source: §A.2 · pp. 14–15, Table 7 · p. 15

A.3 WHY DOES DES WORK?

• Uniformly exploring expert counts in beam search significantly improves the proportion of correct generated answers over vanilla beam search.The ablation distributes rollouts across expert counts from 4 to 11 and omits Diversity-Aware Selection and Experts Configuration Inheritance.Source: §A.3 · p. 15
• Expert Configuration Inheritance filters low-scoring configurations across steps, concentrating later computation on configurations that remain promising.The authors use pass@N to assess whether the final candidate set contains a correct answer.Source: §A.3 · p. 15

B BROADER IMPACT

• The broader-impact discussion presents DES as a way to improve small-model reasoning without retraining or modifying model parameters.It contrasts inference-time computation with the training costs associated with scaling model size.Source: B BROADER IMPACT · pp. 15–16
• The authors describe local deployment on memory-constrained devices as a potential route to more accessible, offline AI assistance.The discussion frames this as potential social impact, including reduced dependence on cloud infrastructure and network connectivity.Source: B BROADER IMPACT · p. 16

C.1 PROMPTS

• Task prompts differ by domain: math prompts request concise or stepwise solutions, code prompts request Python programs, and knowledge prompts request stepwise answers.Math outputs conclude with a boxed answer; knowledge answers are prompted in a solution-tag format.Source: §C.1 · pp. 16–17

C.2 INFERENCE DETAILS

• Inference uses up to 10 steps, temperature 0.8, eight initial expert counts from 4 to 11, and four NVIDIA RTX A6000 GPUs.The candidate count retained per step is set to N/4, where N is the predefined total number of final candidates.Source: §C.2 · p. 17
• Policy models are served with vLLM and reward models with Transformers; task-specific post-processing extracts math answers, code blocks, or knowledge answers in solution tags.Source: §C.2 · p. 17

D LICENSE

• The paper states that code will be publicly accessible upon acceptance and that community-standard licenses are used.Source: D LICENSE · p. 18

• Qwen3-30B-A3B is listed with Apache, and Ling-lite-1.5 with MIT.Source: §2 · p. 18
• OLMoE-1B-7B-Instruct is listed with Apache, and DeepSeek-V2-Lite-Chat with Deepseek.Source: §2 · p. 18
• Llama3.1-8B-PRM-Deepseek-Data is listed with Llama, and Qwen2.5-Math-PRM-7B with Qwen.Source: §2 · p. 18