Research map / Large language models / Agents, code and math
Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Model Swarms, a PSO-inspired collaborative search that moves a pool of LoRA-tuned Gemma-7B experts in weight space to optimize a black-box utility function, adapting them with as few as 200 examples.
Problem
Adapting and composing diverse LLM experts for new tasks, domains, reward models, or human interests without large tuning sets or strong assumptions about how experts should be combined.
Approach
Treat each LLM expert as a particle in weight space and update velocities using personal best, global best, and global worst (repel) terms under a model-to-scalar utility function, with pairwise interpolation for initialization, restarts, and step-length decay; the global best is returned as the adapted expert.
Key results
- Model Swarms outperforms 12 composition baselines by 13.3% on average across 9 single-task datasets, with 21.0% gain on reasoning tasks.
- On reward models, Model Swarms beats all 14 baselines (including PPO and DPO) by 6.7% on average, and achieves the best verbose and concise RM scores.
- In human interest evaluation, Model Swarms achieves a 70.8% average human win rate versus initial experts, with LLM-judge scores improving 17.6% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use collaborative Model Swarms search instead of the strongest baseline | Gemma-7B experts · GSM8K test | score: .354 -> .459 | Helped |
| Use collaborative Model Swarms search instead of DPO | Gemma-7B experts · Verbose RM test | reward score: .682 -> .770 | Helped |
| Add random interpolation to expand the initial expert pool (ablation) | Model Swarms · MMLU | score: 0.504 -> 0.583 | Helped |
| Add random interpolation to expand the initial expert pool (ablation) | Model Swarms · Hellaswag | score: 0.587 -> 0.652 | Helped |
| Add random interpolation to expand the initial expert pool (ablation) | Model Swarms · NLGraph | score: 0.530 -> 0.672 | Helped |
| Add random interpolation to expand the initial expert pool (ablation) | Model Swarms · AbstainQA | score: 0.099 -> 0.175 | Helped |
| Add random starting velocity (ablation) | Model Swarms · MMLU | score: 0.516 -> 0.583 | Helped |
| Add randomness to velocity update weights (ablation) | Model Swarms · MMLU | score: 0.544 -> 0.583 | Helped |
| Add randomness to interpolation, starting velocity, and velocity updates (ablation) | Model Swarms · MMLU | score: 0.528 -> 0.583 | Helped |
| Add crossover through interpolation and initial-pool expansion (ablation) | Model Swarms · MMLU | score: 0.504 -> 0.583 | Helped |
| Use random initial velocity rather than zero velocity (ablation) | Model Swarms · MMLU | score: 0.516 -> 0.583 | Helped |
| Add the global-worst repel term (ablation) | Model Swarms · MMLU | score: 0.534 -> 0.583 | Helped |
Surprising findings
- 89.6% of best-found ending experts did not start as the best expert, and 56.9% started in the bottom half.
- Swarms of only the bottom-half starting experts outperformed the top-1 expert on 2 of 3 datasets.
Limitations
- Requires maintaining and updating all experts each iteration, which is computationally costly and needs homogeneous architectures for weight swarms.
- Not reliable for memorizing new information; perplexity as utility could not be optimized well.
- Dual-use risk: malicious utility functions could be optimized.
Experimental setup
Datasets: Tulu-v2 MMLU MMLU-pro Hellaswag GSM8k Knowledge Crosswords NLGraph TruthfulQA RealToxicityPrompts AbstainQA MedQA MedMCQA
Benchmarks: MMLU MMLU-pro Hellaswag GSM8k Knowledge Crosswords NLGraph TruthfulQA RealToxicityPrompts AbstainQA MedQA MedMCQA LegalBench
Metrics: accuracy reward model score LLM-as-judge score factuality (Facts&Evidence) human win rate correctness surge/emergence
Baselines: Best Single Data Merge Prediction Merge Uniform Soup Slerp Dare-Ties Model Stocks Greedy Soup Pack of LLMs cBTM EvolMerge LoraHub
Builds on: particle swarm optimization LoRA model soups LoraHub EvolMerge
Closest papers by meaning
- Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems (NeurIPS 2025 · similarity 0.91)
- Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.88)
- Multi-Agent Teams Hold Experts Back (ICML 2026 · similarity 0.87)
- AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.87)
- AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems (NeurIPS 2025 · similarity 0.87)
- FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.