RLSVR: Self-Improving AI for Open-Ended Tasks

RLSVR extends RLVR to open-ended tasks by transforming them into verifiable proxy games (e.g., SpyRL). This enables self-play without human preferences, showing gains on summarization and creative writing. It offers a scalable path to self-improvement for LLMs, potentially reducing reliance on expensive reward models. The approach uses information asymmetry to create verifiable proxies.