The authors define ease as the k-NN overlap between two encoders on a small probe set, and ceiling as the average overlap across models drawn from the same training paradigms.
A simple diagnostic predicts whether two AI models will work well together before costly combination
COMPASS measures local neighborhood overlap and paradigm compatibility to rank candidate model pairs for alignment or merging
Top University
Fabian Gröger · Petar Damjanović · Maria Brbić
EPFL · University of Basel · HSLU
Research Digest··2 min read
The authors propose COMPASS, a theoretically motivated score that estimates model-pair compatibility using two cheap-to-compute axes: ease (k-NN overlap) and ceiling (maximum overlap allowed by training paradigms).
Why this paper
From EPFL and 2 others
In one line
Model pair compatibility can be predicted with local neighborhood overlap and paradigm ceiling before alignment or merging.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks (3 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§