TestTime RLVR: Test-Time Reinforcement Learning with Verification and Reasoning

Based on Absolute Zero Reasoner (AZR) Methodology

🎉 News • 🔗 Links • 📝 Roadmap • ⚙️ Algorithm Flow • 📊 Results

✨ Getting Started • 🏋️ Training • 🔧 Usage • 📃 Evaluation

🎈 Citation • 🌻 Acknowledgement • 📧 Contact • 📈 Star History

🚀 TestTime RLVR Implementation

📋 Overview

TestTime RLVR implements test-time reinforcement learning for enhanced reasoning capabilities using the AZR (Absolute Zero Reasoner) methodology. The system generates Input-Program-Output (IPO) triples from benchmark problems and creates three types of reasoning tasks (induction, deduction, abduction) to improve model performance at test time.

🎯 Key Features

Complete Pipeline: LLM Solution Generation → IPO Extraction → Task Generation → LLM Evaluation → Reward Computation
AZR Integration: Full integration with Absolute Zero Reasoner templates and evaluation methods
Benchmark Support: MBPP+ and HumanEval+ datasets with structured data extraction
Execution-based Evaluation: Program execution comparison instead of string matching
VLLM Optimization: Faster inference with VLLM backend support

📈 Implementation Status

✅ Phase 1: Infrastructure Setup - Complete pipeline architecture
✅ Phase 2: Benchmark System - MBPP+/HumanEval+ integration
✅ Phase 3: AZR Template Integration - Three reasoning tasks implementation
✅ Phase 4: Complete Pipeline - Fully functional end-to-end system
🔄 Phase 5: RLVR Training - Reinforcement learning integration (In Progress)

📦 Dataset Setup

Download required benchmark datasets:

# Download MBPP+ and HumanEval+ datasets
wget -O evaluation/code_eval/data/MbppPlus.jsonl https://huggingface.co/datasets/evalplus/mbppplus/resolve/main/MbppPlus.jsonl
wget -O evaluation/code_eval/data/HumanEvalPlus.jsonl https://huggingface.co/datasets/evalplus/humanevalplus/resolve/main/HumanEvalPlus.jsonl

🚀 Quick Start

Running the Pipeline

# Navigate to test directory
cd test/

# Set GPU device
export CUDA_VISIBLE_DEVICES=6

# Execute complete pipeline
bash run_testtime_gpu6.sh

Command Line Options

# From test/ directory
python test_complete_pipeline.py \
    --model "Qwen/Qwen2.5-7B" \
    --benchmark "mbpp" \
    --problem_id "Mbpp/478" \
    --max_tokens 2048 \
    --gpu 6 \
    --verbose \
    --output_dir ../tmp

Batch Evaluation

# From test/ directory
bash run_batch_evaluation.sh "Qwen/Qwen2.5-7B" "mbpp" 10 6

Supported Benchmarks

MBPP+: --benchmark mbpp --problem_id "Mbpp/X"
HumanEval+: --benchmark humaneval --problem_id "HumanEval/X"
Test Mode: --benchmark test (example problems)

📊 Results Structure

tmp/{benchmark}/{problem_id}/          # Single problem results
├── initial_solution/                  # LLM's original solution + correctness
│   ├── {problem_id}_original_problem.txt    # Original benchmark problem
│   ├── {problem_id}_llm_solution.txt        # LLM solution + correctness evaluation
│   └── {problem_id}_extracted_program.py   # Extracted function code
├── ipo_triples/                      # Input-Program-Output triples  
├── task_prompts/                     # Generated reasoning tasks
├── llm_responses/                    # LLM responses to tasks
├── extracted_answers/                # Extracted answers from responses
├── {problem_id}_reward_analysis.json
├── {problem_id}_reward_summary.txt
└── {problem_id}_pipeline_summary.json

test/batch_results/                   # Batch evaluation results
├── batch_evaluation_{timestamp}/
│   ├── batch_evaluation_results.json      # Detailed results with correctness stats
│   └── evaluation_summary.md              # Summary report with accuracy rates

🔗 AZR References

📝 TestTime RLVR Roadmap

✅ Complete Pipeline Implementation

✅ IPO Triple Extraction with Structured Data

✅ Three Reasoning Tasks (Induction/Deduction/Abduction)

✅ Execution-based Evaluation System

🔄 VeRL Integration for RLVR Training

📋 Multi-Problem Batch Processing

⚙️ TestTime RLVR Algorithm Flow

TestTime RLVR implements a comprehensive test-time reasoning pipeline based on AZR methodology:

🔄 Pipeline Stages

LLM Solution Generation: The model generates an initial solution for a given benchmark problem (MBPP+/HumanEval+)
IPO Triple Extraction: Input-Program-Output triples are created using structured benchmark data and LLM solution execution
Task Generation: Three types of reasoning tasks are generated:
- Induction: Deduce function from input/output pairs + message
- Deduction: Predict output from code + input
- Abduction: Predict input from code + output
LLM Evaluation: The model attempts to solve the generated reasoning tasks using AZR prompts and templates
Reward Computation: Solutions are verified through program execution, receiving accuracy-based rewards

🎯 Key Innovations

Structured Data Integration: Direct use of benchmark base_input/plus_input instead of assert parsing
Execution-based Evaluation: Program execution comparison for accurate task evaluation
Function Name Normalization: Consistent f function naming following AZR methodology
Docstring Utilization: LLM-generated docstrings enhance induction task quality

📊 Results

Main Results

Our approach achieves strong performance across both code and math reasoning benchmarks without using any external data:

Model	Base	#data	Code Avg	Math Avg	Total Avg
Base Models
Qwen2.5-7B	-	-	52.0	27.5	39.8
Qwen2.5-7B-Ins	-	-	56.3	37.0	46.7
Qwen2.5-7B-Coder	-	-	56.6	23.9	40.2
Reasoners Trained on Curated Code Data
AceCoder-RM	Ins	22k	58.3	37.4	47.9
AceCoder-RM	Coder	22k	57.3	27.5	42.4
AceCoder-Rule	Ins	22k	55.4	36.9	46.2
AceCoder-Rule	Coder	22k	60.0	28.5	44.3
CodeR1-LC2k	Ins	2k	60.5	35.6	48.0
CodeR1-12k	Ins	10k	61.3	33.5	47.4
Reasoners Trained on Curated Math Data
PRIME-Zero	Coder	484k	37.2	45.8	41.5
SimpleRL-Zoo	Base	8.5k	54.0	38.5	46.3
Oat-Zero	Math	8.5k	45.4	44.3	44.9
ORZ	Base	57k	55.6	41.6	48.6
Absolute Zero Training w/ No Curated Data (Ours)
AZR (Ours)	Base	0	55.2 +3.2	38.4 +10.9	46.8 +7.0
AZR (Ours)	Coder	0	61.6 +5.0	39.1 +15.2	50.4 +10.2

Scaling Results

AZR shows consistent improvements across model sizes and types:

Model Family	Variant	Code Avg	Math Avg	Total Avg
Llama3.1-8b		28.5	3.4	16.0
Llama3.1-8b	+ AZR (Ours)	31.6 +3.1	6.8 +3.4	19.2 +3.2
Qwen2.5-3B Coder		51.2	18.8	35.0
Qwen2.5-3B Coder	+ AZR (Ours)	54.9 +3.7	26.5 +7.7	40.7 +5.7
Qwen2.5-7B Coder		56.6	23.9	40.2
Qwen2.5-7B Coder	+ AZR (Ours)	61.6 +5.0	39.1 +15.2	50.4 +10.2
Qwen2.5-14B Coder		60.0	20.2	40.1
Qwen2.5-14B Coder	+ AZR (Ours)	63.6 +3.6	43.0 +22.8	53.3 +13.2

✨ Getting Started

🎄 Environment Setup

conda env create -f azr_env.yml
conda activate azr
pip install -r flashattn_requirements.txt

💾 Data Processing

Process evaluation data on CruxEval / LiveCodeBench Execution during AZR Self-play

python -m absolute_zero_reasoner.data_construction.process_code_reasoning_data

🏋️ Training

⚠️WARNING⚠️: The Python executor in this repository is very raw and intended for research purposes only. It is not secure for production environments. We plan to update our executor to more secure implementations in the future. Your use of our code is at your own discretion and risk.

🫛 Seeding (Optional)

We provide the seed datasets we collected by prompting each model in data/. If you want to create your own seed data, use the following script:

export OUTPUT_SEED_PATH=data/<new_ded_abd_seed_data_name>.jsonl
export OUTPUT_CODE_F_SEED_PATH=data/<new_ind_seed_data_name>.jsonl
bash scripts/seeding/<7b|14b|coder3b|coder7b|coder14b|llama>.sh

♟️ Self-play

3b models need 2 X 80gb GPUs, 7/8b models need 4 X 80gb, 14b requires 8 X 80gb

bash scripts/selfplay/<7b|14b|coder3b|coder7b|coder14b|llama>.sh

If you want to use your own ded/abd or ind seed dataset:

export OUTPUT_SEED_PATH=data/<your_ded_abd_seed_data_name>.jsonl
export OUTPUT_CODE_F_SEED_PATH=data/<your_ind_seed_data_name>.jsonl
bash scripts/selfplay/<7b|14b|coder3b|coder7b|coder14b|llama>.sh

For using the newly supported sandbox-fusion executor, use docker and set azr.executor=sandboxfusion.

🌚 Resuming Runs

When resuming runs, put the original run wandb id into the script, i.e., trainer.wandb_run_id=<run_id>.

🤗 Converting veRL checkpoints to HF format

python -m absolute_zero_reasoner.utils.convert2hf \
  <veRL_ckpt_path>/actor \
  <veRL_ckpt_path>/actor/huggingface/ \
  <hf_ckpt_path>

📈Design Your Own Intrinsic Rewards!

In configs, just add your own rewards to azr.reward.generation_reward_config, check the ones already implemented such as diversity and complexity rewards. Be Creative!

🔧 Usage

We use the Deepseek R1 & tags as prompt template:

A conversation between User and Assistant. The user asks a question, and the Assistant solves it. The assistant first thinks about the reasoning process in the mind and then provides the user with the answer. The reasoning process and answer are enclosed within <think> </think> and <answer> </answer> tags, respectively, i.e., <think> reasoning process here </think> <answer> answer here </answer>. User: {question}\nAssistant: <think>

📃 Evaluation Code

LiveCodeBench

Setup: LCB needs to first download the data

git clone https://hf-mirror.com/datasets/livecodebench/code_generation_lite evaluation/code_eval/coding/LiveCodeBench/code_generation_lite

Evaluation:

bash evaluation/code_eval/scripts/run_lcb_gen.sh --model <andrewzh/Absolute_Zero_Reasoner-Coder-3b>

Evalplus

New conda env is neede for evalplus

conda create -n evalplus python=3.11
pip install --upgrade "evalplus[vllm] @ git+https://github.com/evalplus/evalplus@d362e933265c3e7e3df8101c930a89c3c470cd9f"
Evaluation:
```bash
condda activate evalplus
bash evaluation/code_eval/scripts/run_evalplus.sh 0 <humaneval|mbpp> <andrewzh/Absolute_Zero_Reasoner-Coder-3b>

Math

Please refer to evaluation/math_eval/README.md for math evaluation.

🎈 Citation

If you find Absolute Zero Reasoner helpful, please cite us.

@misc{zhao2025absolutezeroreinforcedselfplay,
      title={Absolute Zero: Reinforced Self-play Reasoning with Zero Data}, 
      author={Andrew Zhao and Yiran Wu and Yang Yue and Tong Wu and Quentin Xu and Yang Yue and Matthieu Lin and Shenzhi Wang and Qingyun Wu and Zilong Zheng and Gao Huang},
      year={2025},
      eprint={2505.03335},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2505.03335}, 
}

🌻 Acknowledgement

Our reinforcement learning training codebase is a fork of the veRL framework. For rollouts, we used vLLM. The Python executor components are adapted from the QwQ Repository. Additionally, we borrowed our README structure from PRIME. Many thanks to the authors of these projects for their excellent contributions!

📧 Contact

Feel free to contact Andrew Zhao via email: [email protected]