RepetitionCurse: Measuring and Understanding Router Imbalance in Mixture-of-Experts LLMs under DoS Stress

Ruixuan Huang, Qingyue Wang, Hantao Huang, Yudong Gao, Dong Chen, Shuai Wang, Wei Wang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:47750-47764, 2026.

Abstract

Mixture-of-Experts architectures have become the standard for efficient LLM scaling, typically employing expert parallelism to distribute experts across devices. However, the absence of explicit load balancing constraints during inference allows adversarial inputs to trigger severe routing concentration. We demonstrate that out-of-distribution prompts can manipulate the routing mechanism such that all tokens are routed to the same set of top-k experts, which creates computational bottlenecks on certain devices while forcing others to idle. This converts an efficiency mechanism into a denial-of-service attack vector, leading to violations of service-level agreements for time-to-first-token (TTFT). We propose RepetitionCurse, a black-box strategy to exploit this vulnerability. By identifying a universal flaw in MoE router behavior, RepetitionCurse constructs attack prompts using simple repetitive token patterns in a model-agnostic manner. On widely deployed MoE models hosted on 8-GPU clusters, our method increases TTFT by 20% to 148%, significantly degrading service quality.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-huang26bc, title = {{R}epetition{C}urse: Measuring and Understanding Router Imbalance in Mixture-of-Experts {LLM}s under {D}o{S} Stress}, author = {Huang, Ruixuan and Wang, Qingyue and Huang, Hantao and Gao, Yudong and Chen, Dong and Wang, Shuai and Wang, Wei}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {47750--47764}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/huang26bc/huang26bc.pdf}, url = {https://proceedings.mlr.press/v306/huang26bc.html}, abstract = {Mixture-of-Experts architectures have become the standard for efficient LLM scaling, typically employing expert parallelism to distribute experts across devices. However, the absence of explicit load balancing constraints during inference allows adversarial inputs to trigger severe routing concentration. We demonstrate that out-of-distribution prompts can manipulate the routing mechanism such that all tokens are routed to the same set of top-k experts, which creates computational bottlenecks on certain devices while forcing others to idle. This converts an efficiency mechanism into a denial-of-service attack vector, leading to violations of service-level agreements for time-to-first-token (TTFT). We propose RepetitionCurse, a black-box strategy to exploit this vulnerability. By identifying a universal flaw in MoE router behavior, RepetitionCurse constructs attack prompts using simple repetitive token patterns in a model-agnostic manner. On widely deployed MoE models hosted on 8-GPU clusters, our method increases TTFT by 20% to 148%, significantly degrading service quality.} }
Endnote
%0 Conference Paper %T RepetitionCurse: Measuring and Understanding Router Imbalance in Mixture-of-Experts LLMs under DoS Stress %A Ruixuan Huang %A Qingyue Wang %A Hantao Huang %A Yudong Gao %A Dong Chen %A Shuai Wang %A Wei Wang %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-huang26bc %I PMLR %P 47750--47764 %U https://proceedings.mlr.press/v306/huang26bc.html %V 306 %X Mixture-of-Experts architectures have become the standard for efficient LLM scaling, typically employing expert parallelism to distribute experts across devices. However, the absence of explicit load balancing constraints during inference allows adversarial inputs to trigger severe routing concentration. We demonstrate that out-of-distribution prompts can manipulate the routing mechanism such that all tokens are routed to the same set of top-k experts, which creates computational bottlenecks on certain devices while forcing others to idle. This converts an efficiency mechanism into a denial-of-service attack vector, leading to violations of service-level agreements for time-to-first-token (TTFT). We propose RepetitionCurse, a black-box strategy to exploit this vulnerability. By identifying a universal flaw in MoE router behavior, RepetitionCurse constructs attack prompts using simple repetitive token patterns in a model-agnostic manner. On widely deployed MoE models hosted on 8-GPU clusters, our method increases TTFT by 20% to 148%, significantly degrading service quality.
APA
Huang, R., Wang, Q., Huang, H., Gao, Y., Chen, D., Wang, S. & Wang, W.. (2026). RepetitionCurse: Measuring and Understanding Router Imbalance in Mixture-of-Experts LLMs under DoS Stress. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:47750-47764 Available from https://proceedings.mlr.press/v306/huang26bc.html.

Related Material