modelco 1.0.0

LLM collaboration

MoCo: A One-Stop Shop for Model Collaboration Research

MoCo is a toolkit for Model Collaboration research, where multiple language models collaborate and complement each other for compositional AI systems.

Quick Start

conda env create -f environment.yml
pip install modelco
conda activate model_collaboration

Run your first model collaboration experiment (if you don't have 3 GPUs, go to model_collaboration/test_config.json and set "gpu_ids": [0], [0,1], or whatever you have; if your GPU is nice, increase batch_size):

python -m model_collaboration.main -c model_collaboration/test_config.json

You will see the outputs and evaluation results in the model_collaboration/logs/ folder.

Supported Methods

MoCo currently supports the following model collaboration algorithms, across API-level, text-level, logit-level, and weight-level collaboration. We provide a sample config for each method in examples/ and please check out docs/user_readme.md for more details about writing configs and the different collaboration methods implemented.

Method	Core Idea	Code	Sample Config	Doc
API: Nudging	one model guides the decoding of another	link	link	link
API: Prompt Routing	prompt an LM to decide which model to use based on model descriptions	link	link	link
API: Switch Generation	multiple LMs take turns to generate parts of the response	link	link	link
API: Trained Router	train an LM to route based on the dev set	link	link	link
API: Graph Routing	train a graph neural network for routing	link	link	link
API: Cascade	use multiple models in a cascade to improve efficiency	link	link	link
API: Mentor Collab	a mentor model guides a smaller student model for generation	link	link	link
API: Co-LLM	train LMs to defer to another model when uncertain	link	link	link
Text: Multiagent Refine	multiple LMs refine each other's answers iteratively	link	link	link
Text: Multiagent Feedback	multiple LMs provide feedback to each other's answers	link	link	link
Text: Knowledge Card	models generate knowledge paragraphs to assist each other	link	link	link
Text: LLM Blender	use ranker and fuser LMs to combine multiple answers	link	link	link
Text: Heterogeneous Swarms	optimize a graph of multiple LLMs for collaboration	link	link	link
Text: Majority Vote	majority vote	link	link	link
Text: Structured Interaction	execute a structured interaction protocol among LLMs	link	link	link
Text: Multiagent Finetuning	multiple LLMs critique, debate, and refine via finetuning	link	link	link
Text: BBMAS	blackboard-based collaboration among LLMs	link	link	link
Text: Sparta Alignment	models compete and combat for collective alignment	link	link	link
Text: AggLM	RL to train a solution aggregation model	link	link	link
Logit: Logit Fusion	merge the next-token logits from multiple models	link	link	link
Logit: Logit Contrastive	contrast the logits from best/worst models	link	link	link
Weight: Greedy Soup	iteratively consider adding each model's weights from best to worst	link	link	link
Weight: Dare Ties	the dare-ties model merging algorithm	link	link	link
Weight: Model Swarms	particle swarm optimization for models to search in the weight space	link	link	link
Weight: LoraHub	gradient-free optimization of lora combinations	link	link	link
Weight: ExPO	model weight extrapolation	link	link	link

Please note that MoCo does not aim to be a reproducibility study: we adapt the core ideas behind related papers and employ what works flexibly.

Supported Data

MoCo comes with a lot of evaluation datasets built-in, and you are free to bring your own datasets, or even just generate responses only and take evaluation elsewhere. Essentially, change the task and task_type in the config to use diverse datasets. Check out link for more details.

Contributing to MoCo

We welcome contributions to MoCo!

If you are interested in contributing new model collaboration methods, check out link.

If you are interested in contributing new datasets, check out link.

If you have any suggestions, please open an issue.

Citation

If MoCo is helpful for you, please consider citing:

@article{feng2025one,
  title={When one llm drools, multi-llm collaboration rules},
  author={Feng, Shangbin and Ding, Wenxuan and Liu, Alisa and Wang, Zifeng and Shi, Weijia and Wang, Yike and Shen, Zejiang and Han, Xiaochuang and Lang, Hunter and Lee, Chen-Yu and others},
  journal={arXiv preprint arXiv:2502.04506},
  year={2025}
}

// MoCo reference pending

Also, please cite the related papers for the methods you employed, as listed in docs/user_readme.md.

Have a nice day.