LeapLabTHU/Absolute-Zero-Reasoner

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

1893 stars · 298 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

Official implementation of Absolute Zero Reasoner (AZR), a system that trains LLM reasoning via reinforced self-play with zero external data, using a code executor to verify self-generated tasks. It provides training and evaluation pipelines for deduction, abduction, and induction reasoning modes.

Use cases

• train a reasoning model without human-curated datasets
• replicate the Absolute Zero paper results
• run RL with verifiable rewards on code-based reasoning tasks
• evaluate LLMs on math and coding reasoning benchmarks
• experiment with self-play task proposal and solving
• use sandboxed code execution as a reward signal

When to choose

• you want to train or fine-tune LLMs with zero external reasoning data
• you need a research-grade RLVR training pipeline with code-executor verification
• you want to reproduce or extend the Absolute Zero paper

When to avoid

• you need a production-ready inference or serving framework
• you lack GPU resources for RL training
• you want a plug-and-play chatbot rather than a research training codebase

Facets

library · maturity active

llm-training reinforcement-learning machine-learning benchmarking large-language-models machine-learning deep-learning artificial-intelligence python self-play rlvr zero-data reasoning code-executor research-code verifiable-rewards gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="LeapLabTHU/Absolute-Zero-Reasoner")