Research map / Large language models / Agents, code and math
R-Zero: Self-Evolving Reasoning LLM from Zero Data
Chengsong Huang, Wenhao Yu, Xiaoyang Wang, Hongming Zhang, Zongxia Li, Ruosen Li, Jiaxin Huang, Haitao Mi, Dong Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train the Challenger with RL instead of using an untrained Challenger (ablation) | R-Zero · Math average (Qwen3-4B-Base) | average accuracy: 45.01 -> 49.93 | Improved |
| add repetition penalty to the Challenger reward (ablation) | R-Zero · Math AVG (Qwen3-4B-Base) | Math AVG: 45.76 -> 49.07 | Improved |
| add difficulty-based task filtering for Solver data (ablation) | R-Zero · General AVG (Qwen3-4B-Base) | General AVG: 26.69 -> 31.15 | Improved |
| add task filtering for Solver data (ablation) | R-Zero · Math AVG (Qwen3-4B-Base) | Math AVG: 47.35 -> 49.07 | Improved |
| separate Challenger and Solver into two independent models rather than one shared model (ablation) | R-Zero · Math average (Qwen3-4B-Base) | peak performance: Single-R-Zero 47.31 (peak) -> 49.12 | Improved |
| separate Challenger and Solver models (ablation) | R-Zero · pseudo-label accuracy, iteration 1 | pseudo-label accuracy: 63.4% -> 71.0% | Improved |
| run more co-evolution iterations (ablation) | R-Zero · Math average, 0.6B to 4B Qwen3 models | math average | Hurt |
| fine-tune R-Zero checkpoint on human-labeled data afterwards | R-Zero · Math average (Qwen3-4B-Base) | average accuracy: +2.35 points over human-label-only baseline | Improved |
| mix human-labeled data into the R-Zero training data instead of sequential SFT (ablation) | R-Zero · Qwen3-4B-Base math and general | accuracy | Mixed |
| pseudo-label accuracy declines as the Challenger makes harder questions | R-Zero · GPT-4o-checked question sets (Qwen3-4B-Base) | pseudo-label accuracy: 79.0% -> 63.0% | Hurt |
| remove the math-only constraint from the Challenger prompt (ablation) | R-Zero · general benchmarks (Qwen3-8B) | SuperGPQA: 31.38 -> 32.29 | Improved |
| remove the math-only constraint from the Challenger prompt (ablation) | R-Zero · math and general benchmarks (Qwen3-4B) | AMC: 57.27 -> 56.27 | No clear effect |
Later papers that compared with this method
4 later accepted papers report R-Zero as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- R-Diverse: Mitigating Diversity Illusion in Self-Play LLM Training (ICML 2026 · better on 8 of 8 shared results)
- G-Zero: Self-Play for Open-Ended Generation from Zero Data (NeurIPS 2026 · better on 6 of 8 shared results)
- D²Evo: Dual Difficulty-Aware Self-Evolution for Data-Efficient Reinforcement Learning (ICML 2026 · better on 2 of 2 shared results)
- Agentic Proposing: Enhancing Large language Model Reasoning via Compositional Skill Synthesis (ICML 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Absolute Zero: Reinforced Self-play Reasoning with Zero Data (NeurIPS 2025 · similarity 0.91)
- General-Reasoner: Advancing LLM Reasoning Across All Domains (NeurIPS 2025 · similarity 0.90)
- RL of Thoughts: Navigating LLM Reasoning with Inference-time Reinforcement Learning (ICLR 2026 · similarity 0.90)
- Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.89)
- Diversity-Aware Policy Optimization for Large Language Model Reasoning (NeurIPS 2025 · similarity 0.89)
- Tina: Tiny Reasoning Models via LoRA (ICLR 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
