Resonant Weight Grids
The proposal: store model weights in a new format so two agents' skills can be merged into one model without retraining. This page tests the core idea on a toy. Each "model" is a list of 48 numbers; a skill is the change a fine-tune made to some of those numbers. You merge two skills and see how much of each survives.
Claimed by the proposal, not verified
- Weights encoded as "persistent homology signatures" with lossless functional equivalence. No published method does this; persistent homology summarizes shape and throws detail away, so "lossless" is very unlikely.
- 98% capability retention after a "harmonic merge" versus 62% for parameter averaging.
- Merging works across different models and adapts dimensionality to available compute.
- Released open source with reference implementations and a benchmark suite. No link or repository was given.
| Merge method | Skill A kept | A | Skill B kept | B |
|---|
"Kept" = how much of each skill's change still points the same way and has the same size after the merge (projection of the merged change onto that skill's change, capped at 100%). Real models would need task benchmarks, not this score.
Strip above: each column is one of the 48 numbers. Top row skill A, middle skill B, bottom the sign-elect merge. Blue = pushed up, amber = pushed down.
What this shows, and what a real build needs
Merging fine-tunes without retraining is real and well studied (task arithmetic, TIES, DARE, model soups). It works when both parents are fine-tunes of the same base model and their changes don't fight. Switch "different models" on and every method falls apart, because the numbers no longer mean the same thing in both parents. That is the part the proposal claims to have solved, and nothing here supports it.
A real test for the fleet would be: take one open base model, make two LoRA fine-tunes, merge them with an existing tool such as mergekit, and score both tasks before and after. A new weight format is not needed for that.
What to try
- Set overlap to 0. Plain averaging keeps about 50% of each skill; adding the changes keeps 100%. That gap is where the 62% versus 98% style numbers come from.
- Raise overlap and conflict to 80%+. Watch adding fall apart and sign-elect hold up better but still lose ground.
- Switch to "different models". All methods drop toward noise. That is the cross-model claim.
- Push trim to 70%. Small but real parts of a skill get cut, so retention drops even with no conflict.