Mechanism

Operates on task vectors τm=θm−θbase\tau_m = \theta_m - \theta_{\text{base}}:

  • Task Arithmetic: θnew=θbase+∑mλmτm\theta_{\text{new}} = \theta_{\text{base}} + \sum_m \lambda_m \tau_m.
  • TIES-Merging (Trim, Elect Sign, Merge): Resolves parameter interference across multiple task vectors by (1) trimming the smallest 1−k%1-k\% magnitude deltas to zero, (2) electing the dominant sign sgn(∑mτm)\text{sgn}(\sum_m \tau_m) per coordinate, and (3) averaging only coordinates matching the elected sign.
  • DARE (Drop And REscale): Randomly drops p∈[0.9,0.99]p \in [0.9, 0.99] of fine-tuned delta parameters τm\tau_m using a Bernoulli mask and rescales the surviving 1−p1-p deltas by 11−p\frac{1}{1-p} before merging, eliminating cross-specialist interference.

References

  1. Editing Models with Task Arithmetic ↗

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi et al. · 2022

    All 7 authors

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, Ali Farhadi

  2. TIES-Merging: Resolving Interference When Merging Models ↗

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, Mohit Bansal · 2023

  3. Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch ↗

    Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, Yongbin Li · 2023