# Absolute Zero > Self-play RL paradigm where the model proposes its own training tasks using code execution as a verifier, requiring zero external data - Authors: Ben - Published: 2025-05-09 - Source: Bevaya Labs, --- # Absolute Zero: Reinforced Self-play Reasoning with Zero Data **Paper**: [Absolute Zero: Reinforced Self-play Reasoning with Zero Data](https://arxiv.org/abs/2505.03335) **Presenter**: Yosh **Note Taker**: Ben **Date**: May 9, 2025 --- ## Overview Introduces a new RLVR paradigm where a model autonomously proposes and solves its own reasoning tasks without any external training data. The Absolute Zero Reasoner (AZR) uses code execution to both validate proposed tasks and verify solution correctness, creating a unified verifiable reward signal. Despite training entirely without human-curated examples, AZR surpasses zero-setting models trained on tens of thousands of in-domain examples on coding and mathematical reasoning benchmarks. --- ## Novelty - Task self-proposal: the model generates its own training problems, grounded by code execution to ensure tasks are well-posed and verifiable. - Unified reward from code execution: the same execution environment validates both task correctness (is this a valid problem?) and solution correctness (is this answer right?). - Zero external data: the entire training loop operates without human-curated question-answer pairs. - Self-generated curriculum: task difficulty emerges organically from the model's evolving capabilities rather than from hand-designed difficulty schedules. --- ## Learnings - Code execution as a verifier is a powerful grounding mechanism: it provides reliable binary rewards without human annotation or an LLM judge. - Self-proposed tasks appear to form an effective curriculum—the model naturally gravitates toward problems at the frontier of its current ability. - Zero-data RL can surpass supervised approaches trained on large curated datasets, challenging the assumption that human-labeled data is necessary for strong reasoning. - Results hold across model scales and classes, suggesting the paradigm is broadly applicable. --- ## Notes for Application - Code execution as a reward signal is directly applicable to structured extraction tasks where outputs can be programmatically validated (e.g., JSON schema checks, field format verification). - The self-proposal loop could be adapted for document understanding: a model generates candidate extraction queries, attempts to answer them, and receives rewards from a deterministic validator. - Zero-data self-play reduces reliance on annotated training sets, which could help in domains where labeled data is scarce (e.g., rare form types or new document categories). - The grounding requirement (tasks must be verifiable by execution) keeps self-play from drifting into unverifiable territory—an important constraint to preserve when adapting the approach.