Absolute Zero: Reinforced Self-play Reasoning with Zero Data
What this work does
One language model proposes and solves executable reasoning tasks, using validity and learnability signals to generate its own curriculum.
Connection to discovery
A concrete mechanism for collecting new verifiable tasks without a prewritten question-answer corpus.
Evidence
The body states deterministic-program checks, task buffers and distinct verification rules for the three task modes.
Scope and limitations
The executable language, filters, task modes and pretrained prior remain supplied. Abductive program completion is not automatically identification of a real-world causal mechanism.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.