← All questions
Training & Fine-tuningHard· Gumbel· Reparameterization· Sampling

Why use Gumbel noise specifically for retriever trainin...

The max-stability property of Gumbel noise allows the perturbation to act as a reparameterization rather than a heuristic smoothing. While Gaussian noise is intuitive, it results in the probit model, which is mathematically intractable for multi-class choice probabilities. Bias in this setup comes from the temperature parameter and straight-through substitution, not the noise itself.

Code playground (Python)