New Trick Makes AI Merging Faster and More Accurate
This could make AI models cheaper to update and more reliable for everyday tasks.
Model merging combines fine-tuned models by adding their task vectors to a shared backbone, but interference is usually assessed only after building and evaluating a candidate merge, making coefficient search expensive. The article reports that a first-order directional derivative can predict interference only when evaluated at the appropriate model.
At the receiver model, it forecasts nearby merges per task and per example, halves the accuracy error of a no-change baseline, and identifies most examples whose predictions flip. For full multi-task merges, a single evaluation is insufficient; two evaluations along the merge path are needed.
At the shared pretrained backbone, the derivative was nearly orthogonal to realized changes in vision experiments and only partly predictive in LoRA language-model pilots. The authors use frozen receiver-local responses to construct LR-Merge, a label-free correction that improves all tested closed-form, isotropic, and sparsified base mergers, with gains generally increasing as more tasks are merged. Once built, the surrogate answers new coefficient and task-subset queries in seconds.
- Merging AI models often causes interference, but a new method predicts it before merging.
- LR-Merge fixes interference automatically, improving accuracy and speed.
- It works best when merging similar models and may not suit all cases.
Why It Matters
Faster, cheaper AI updates could lead to more capable and personalized tools for work and daily life.