Reproduce “Gatekeeper: Improving Model Cascades Through Confidence Tuning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Stephan Rabanser, Nathalie Rauschmayr, Achin Kulshrestha, Petra Poklukar, Wittawat Jitkrittum, Sean Augenstein, Congchao Wang, Federico Tombari
Gatekeeper fine-tunes a small model so its confidence better indicates correctness in a cascade that can defer requests to a larger model. It reinforces correct predictions and lowers confidence on incorrect ones. In a Gemma2B-to-Gemma7B cascade, deferral performance is 7× and 10× higher than an untuned baseline on ARC-e and ARC-c, respectively; small-model accuracy can decline.
Key insights
- Explicitly penalizing overconfidence in incorrect predictions, while reinforcing confidence in correct ones, significantly improves the small model's ability to discern what it knows and doesn't know.
- The tunable α hyperparameter provides a direct control mechanism for balancing M_S's overall accuracy against its effectiveness as a deferral agent within the cascade.
- The method is architecture-agnostic, demonstrating improved confidence calibration and deferral performance across encoder-only, decoder-only, and encoder-decoder model types without architectural changes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.