To Intervene or Not: Guiding Inference-time Alignment with Probabilistic Model Blending

In three linesBlendIn, an inference-time alignment framework, probabilistically blends distributions from two models instead of applying binary guidance. The method weights each model's contribution based on reliability, achieving up to 50% performance improvement on challenging model pairs.

Summary generated by Claude — human-verified