TL;DR

This work frames task-distribution design as an optimization problem and introduces PAIRED, where an environment generator uses regret to create solvable but challenging curricula.

Why it matters

When manual environment design becomes the bottleneck, automatic curricula offer a path to scaling useful experience with the learner.

Key findings

  1. 01

    Regret can guide environment generation toward tasks that are neither trivial nor impossible.

  2. 02

    The training distribution can adapt as agent capability changes.

Scaling dimensions

Models, methods & benchmarks

Algorithms
PAIRED, Unsupervised environment design

Topics

Read the original sourcearXiv