Reinforced self-play with zero external data

Absolute Zero Reasoner

A browser-based simulation of a single model that proposes its own code reasoning tasks, verifies them with an executor, solves them, and uses the resulting rewards to build an increasingly challenging curriculum.

Educational AZR Simulation

Self-Evolving Training Loop

01Propose

Create a task near the solver’s current capability frontier.

02Validate

Execute the generated program and reject invalid tasks.

03Solve

Use the same model to reason over the self-generated task.

04Verify

Compare the answer with executable ground truth.

05Learn

Update proposer and solver behavior from verifiable rewards.

Current Self-Generated Task

Deduction
Difficulty 0.00
Press Start or Single Step to let the model propose its first task.
Model answer
—
Executor ground truth
—

Verifiable Reward

0.00combined reward
Task validity—
Learnability reward—
Solver accuracy reward—
Predicted success—

The learnability score is highest when predicted success is near 50%, representing tasks that are neither trivial nor impossible.

Capability Growth

Deduction Abduction Induction Rolling accuracy

Training Trace

This simulation illustrates the paper’s concepts; it does not reproduce the authors’ neural-network training, TRR++, or benchmark results.