Many fine-tunes, one set of weights.

The problem, plainly

Two fine-tunes of one base each move millions of parameters. Most moves are tiny and carry little. A few are large and carry the skill. When the two fine-tunes move the same parameter in opposite directions, a plain average lands near zero, and both skills lose that piece.

Task vectors

Write each fine-tune as base + delta. The delta, or task vector, is what the fine-tune learned. Merging becomes a question about deltas: which to keep, how much of each, and what to do where they disagree.

Trimming

Many entries in a delta are noise. Keeping only the largest fraction of each delta reduces interference between models before anything is combined.

Electing a sign

For each parameter, the merge picks the direction with the most total weight behind it, and keeps only the deltas that agree with it. Updates that disagree are dropped for that parameter rather than averaged into nothing.

Dropping and rescaling

Randomly dropping a share of each delta and rescaling the remainder can preserve a fine-tune's behaviour while leaving more room for others. It is a cheap way to make deltas sparser before they meet.

Why only one base

These methods rely on every model sharing the same parameter layout and starting point. Fine-tunes of different bases do not line up parameter for parameter. Merging across families is what expert transplant is for.