A field guide to frankensteining neural networks

What it does This repository is a curated reading list and taxonomy companion to the 2026 ACM Computing Surveys paper Model Merging in LLMs, MLLMs, and Beyond. It catalogs papers on techniques that combine already-trained neural networks—often without access to original training data or expensive retraining—across methods, theories, and domains.

The interesting bit Rather than a flat list of links, the repository mirrors the survey’s taxonomy: pre-merging fine-tuning tricks, weight-based and subspace-based fusion, dynamic routing, and post-calibration. It also flags which studies ran experiments on models with ≥7B parameters, making it easier to find scale-relevant work.

Key highlights

• Tied to a peer-reviewed ACM Computing Surveys article (2026), giving it an editorial backbone beyond a casual awesome-list.
• Papers are grouped by technique (e.g., sparse subspace merging, routing-based dynamic merging) and by application domain (LLM alignment, multimodal fusion, diffusion style mixing, continual learning, federated learning).
• Explicitly tags experiments conducted on models with ≥7B parameters (or mainstream small LLMs), helping readers filter for scale-relevant work.
• Coverage extends to adjacent topics like model merging for knowledge unlearning, MoE efficiency, deepfake detection, and adversarial attack or defense.
• Accepts pull requests for missing papers and re-categorization, suggesting it is intended as a living document.

Caveats

• This is a bibliography, not a framework: you will find paper titles and links, not runnable code or a merging library inside this repository.
• The README shows visible truncation in places (e.g., the RobustMerge benchmark entry is cut off), indicating the list is still being updated and may contain incomplete entries.

Verdict Researchers and engineers drowning in the recent flood of model-merging preprints will find this a useful navigational chart. If you are hunting for a drop-in Python toolkit to merge your own checkpoints, look elsewhere.