We introduce a method for training process reward models for math reasoning without any supervision, i.e., it requires neither step-level annotations nor ground-truth verification of final answers.

We introduce a joint inference framework for large (vision) language models to perform unsupervised adaptation on a given task, resulting in the improved performance upon independent zero-shot predictions.

An approach to perform fully unsupervised transfer on a downstream dataset given representation spaces of foundation models. Although being fully unsupervised, our approach outperforms CLIP zero-shot transfer and sometimes matches optimal supervised performance!

We develop the approach to infer fine-grained labels given coarsely labeled dataset.

Simple model-agnostic framework for inferring human labeling of a given dataset without any external supervision.

Tractable discrete distributions and stochastic optimization w.r.t. discrete distributions over structured objects (i.e. graphs, permutations, top-k).

Variance reduction techniques for stochastic optimization w.r.t. a distribution over permutations.

The generative model for definitions of polysemous words given vector representation of the word and example of its use.