Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

ElasticMM: Efficient Multimodal LLMs Serving with Elastic Multimodal Parallelism

Authors: Zedong Liu, Shenggan Cheng, Guangming Tan, Yang You, Dingwen Tao

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct a comprehensive evaluation of Elastic MM on two real-world datasets. Compared to the SOTA baseline v LLM [16], Elastic MM reduces TTFT by up to 4.2 and achieves a 3.2 4.5 throughput improvement while meeting service-level objective (SLO) requirements. [...] 4 Evaluation
Researcher Affiliation Academia 1Institute of Computing Technology, Chinese Academy of Sciences 2University of Electronic Science and Technology of China 3National University of Singapore
Pseudocode No The paper describes methods and processes in paragraph form and through mathematical proofs in Appendix B, but it does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures.
Open Source Code No Our experiments use multiple mainstream open-source models and datasets, making them inherently reproducible. We will soon release our code for full reproduction.
Open Datasets Yes Two open-source multimodal datasets are used, each containing a mix of multimodal and text-only requests: Visual Web Instruct[42] is a large-scale dataset collected from over 700K unique web URLs; Share GPT-4o[43] comprises 50K images of varying resolutions along with corresponding text prompts sourced from the multimodal GPT-4o model.
Dataset Splits No Following prior work [16, 36], we use a Poisson distribution to generate variable request arrival rates (requests per second, QPS) and incorporate real-world production service traces to simulate realistic workload distributions. [...] To demonstrate the robustness of these optimizations, we generate requests by sampling from a mixed dataset composed of two distinct sources.
Hardware Specification Yes We evaluate Elastic MM on a high-end workstation equipped with eight NVIDIA A800 80GB GPUs, two 64-core Intel Xeon 8358P CPUs, and 2 TB of DDR4 memory.
Software Dependencies Yes We implement Elastic MM using 7,000 lines of code based on Python, C++. We build upon v LLM [16] (v0.6.6), a stateof-the-art generative model inference platform.
Experiment Setup Yes Following prior work [40, 36], we set the SLO to 10 the latency under light load and then scale it with a constant factor (ranging from one to five) to evaluate performance under both relaxed and strict conditions.