The authors define a model’s “compositional strength” using the Shapley value, a cooperative game-theory measure of each participant’s marginal contribution across possible groups.
A model’s solo performance poorly predicts its value in teams
Across multi-model evaluations, the authors find that directly rewarding system-level outcomes produces more useful collaborators than training for standalone correctness.
Big Tech
Shangbin Feng · Yulia Tsvetkov · Jan Kautz · David Acuna
University of Washington · NVIDIA
Research Digest··3 min read
Feng et al.
Why this paper
From NVIDIA and University of Washington
In one line
A model's contribution to a multi-model system is its compositional strength, uncorrelated with individual performance, and trainable via C-RL.
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 (2 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§