[KongchangAI]
Concept非对称自博弈

Asymmetric Self-Play

A training paradigm in reinforcement learning and multi-agent games in which two agents with asymmetric capabilities or information compete, one posing problems and the other solving them, forming a curriculum with automatic difficulty adjustment.

Timeline (last 90 days)

Nov 6

The Test-Time Self-Play (TTSP) method proposed in the paper bypasses the dependence on verifiers through asymmetric self-play, offering a new idea for test-time scaling without a verifier.

Unverified50%
Nov 6

Asymmetric self-play was first systematized in reinforcement learning and multi-agent game research, with the typical setup being two agents asymmetric in capability or information competing, one responsible for posing problems or setting goals and the other for achieving them, forming an automatic curriculum.

Unverified50%

All Facts (2)

Source Articles