What is model merging, and when would you use it?
Deep learning · Pro 2
Short answer
Model merging combines the weights of models fine-tuned from the same base into one model, by averaging, SLERP, or methods like TIES and DARE that drop small changes first. Use it to join skills trained separately without forgetting, to serve one model instead of several, or to fit on a device. Always test every task after merging.
Deeper answer
The change fine-tuning makes is a task vector: fine-tuned minus base weights. Adding task vectors combines skills; subtracting can remove a behaviour. Simple averaging works surprisingly well, but tasks can interfere; TIES keeps large changes and resolves sign conflicts, DARE drops changes randomly and rescales.
Unlike an ensemble, which runs several models per request, a merge serves one model at normal cost. Merging needs no GPUs for simple methods. Precise tasks may degrade, so keep separate adapters as fallbacks and check safety behaviour too.
What a strong answer covers
- Combines weights of models from the same base
- Names methods: average, SLERP, TIES, DARE
- Use cases: skills without forgetting, one model, devices
- Tests every task and safety after merging
Answer it out loud in the mock interview about 1 min
Follow-ups to expect
- How is merging different from an ensemble?
- What would you do if one task got worse after merging?
Common mistakes
- Merging models from different base models
- Testing only one task after merging
Learn it
Model merging: combining fine-tuned models into one
- Explain backpropagation in simple terms.
- What is the Chinchilla scaling law, and why do labs now train models far beyond it?
- How would you estimate the compute and cost to train a language model from scratch?
- Which fine-tuning settings matter most, and how do you choose them?
- Your neural network overfits. What do you try?
- What is transfer learning, and when would you use it?
- What are vanishing and exploding gradients, and how are they addressed?
- Why do convolutional neural networks work well for images?