EnnengYang/Awesome-Model-Merging-Methods-Theories-Applications

What it is

This is a curated, constantly updated collection of academic papers about model merging — the practice of combining the weights of two or more trained AI models (like LLMs) into one stronger or multi-capable model, without needing the original training data or heavy computation.

It accompanies a specific survey paper (published in ACM Computing Surveys, 2026) and aims to be a living resource for researchers and practitioners.

What it contains

The README organizes hundreds of papers into a clear structure:

• Survey papers — overviews of model merging.
• Benchmarks & evaluation tools — datasets, libraries, and empirical comparisons (e.g., MergeKit, FusionBench).
• Advanced methods — subdivided into pre‑merging steps (fine‑tuning strategies, weight alignment), during‑merging techniques (weighted averaging, subspace methods, routing), and theoretical analyses.
• Applications — across large language models (LLMs), multimodal LLMs (MLLMs), image/video generators, and many classic ML subfields (continual learning, federated learning, adversarial defense, etc.).

Each entry in the tables shows the paper title, year, venue, and often a Remark column listing the models used in experiments (e.g., Llama‑3‑8B, Mistral‑7B). The project also flags papers that used models of size ≥7B, as noted in the News section.

How to use it

• Find relevant papers quickly by browsing the categorized tables.
• Check the “Remark” column to see which models a study used, aiding reproducibility.
• Contribute – the README explicitly invites pull requests to add new papers, fix entries, or suggest better categorization.

It is essentially a structured, browsable bibliography for anyone interested in model merging, from beginners seeking key papers to experts tracking the latest work.