R-Zero: Self-Evolving Reasoning LLM from Zero Data
What this work does
Co-evolves a challenger and solver initialized from a base language model, using solver uncertainty and repetition controls to create a curriculum and self-consistency-derived training targets.
Connection to discovery
An enabling curriculum-generation and self-training mechanism for systems that expand their learning challenges.
Evidence
The body specifies alternating GRPO updates, filtering rules and math/general-reasoning evaluations with iteration and module ablations.
Scope and limitations
Zero external training data does not remove pretrained knowledge; agreement-based targets and reasoning benchmark gains do not independently validate scientific hypotheses.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.