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..

Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems

Authors: Shangbin Feng, Zifeng Wang, Palash Goyal, Yike Wang, Weijia Shi, Huang Xia, Hamid Palangi, Luke Zettlemoyer, Yulia Tsvetkov, Chen-Yu Lee, Tomas Pfister

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that HETEROGENEOUS SWARMS outperforms 17 roleand/or weight-based baselines by 18.5% on average across 12 tasks. Further analysis reveals that HETEROGENEOUS SWARMS discovers multi-LLM systems with heterogeneous model roles and substantial collaborative gains, and benefits from the diversity of language models.
Researcher Affiliation Collaboration 1University of Washington 2Google Cloud AI Research 3Google EMAIL EMAIL
Pseudocode Yes Algorithm 1: Particle Swarm Optimization step (PSO) Input: vectors {xi}n i=1 and the utility values {f(xi)}n i=1 by utility function f Algorithm 2: JFK-score Input: adjacency matrix Abest and models {xi}n i=1 Algorithm 3: Heterogeneous Swarms Input: language models {xi}n i=1, utility function f Algorithm 4: DAG Decoder (G-decode) Input: continuous adjacency matrix A Rn n where aij denotes the likelihood of a directed edge from model xi to model xj
Open Source Code No Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We used open-access and publicly available data (Section D), code will be made publicly available upon acceptance.
Open Datasets Yes We employ 12 datasets to evaluate HETEROGENEOUS SWARMS and baselines spanning knowledge, reasoning, agent, and miscellaneous capabilities. We filter examples in GAIA to retain examples where the humanprovided tool use contexts could support the final answer for GAIA-text. ... MMLU-pro [94], Knowledge Crosswords [19], COM2 [23], GSM8k [13], NLGraph [89], Normad [76], GAIA-text [67], the knowledge graph and lateral thinking puzzle subtasks of Agent Bench [61]; 4) miscellaneous: long-context with Qasper [16], reliability with Abstain QA [25], and LLM-as-a-judge with Wo W [106].
Dataset Splits Yes We by default sample 200 examples for optimization and 1,000 for evaluation, while downsampling if there’s not enough data.
Hardware Specification Yes Experiments are performed on a cluster with 16 A100 GPUs each with 40 GB memory.
Software Dependencies Yes We implement a prototype of HETEROGENEOUS SWARMS with GEMMA-7B (google/gemma-7b-it) [29] in the main paper and also employ other LLMs such as MISTRAL-7B in Table 5. ... (mistralai/Mistral-7B-Instruct-v0.3) [46]
Experiment Setup Yes We employ p = 0.8 for top-p sampling, N = 10, M = 10, search patience 6, max iteration 20, while running grid search over other hyperparameters and report performance of the best-found multi-LLM systems. ... Specifically, ϕv {0.1, 0.2, 0.3}, ϕp {0.1, 0.2, 0.3, 0.4, 0.5}, ϕg {0.2, 0.3, 0.4, 0.5, 0.6}, ϕw {0.01, 0.05, 0.1}, λ {0.5, 0.6, 0.7, 0.8, 0.9, 1.0}.