JustRL demonstrates that competitive reinforcement learning performance for small language models doesn't require complex multi-stage pipelines or dynamic schedules. Using a minimal recipe with single-stage training and fixed hyperparameters, we achieve state-of-the-art results on mathematical reasoning tasks. This repository contains a lightweight evaluation script to reproduce evaluation results for JustRL models on nine challenging math benchmarks.

We release two models:

• JustRL-DeepSeek-1.5B: Trained from DeepSeek-R1-Distill-Qwen-1.5B
• JustRL-Nemotron-1.5B: Trained from OpenMath-Nemotron-1.5B

Both models use identical hyperparameters without per-model tuning, demonstrating the robustness of our approach.

✨ Simplicity: Single-stage training with fixed hyperparameters, without multi-stage pipelines or dynamic schedules

📈 Stability: Smooth, monotonic improvement over 4,000+ training steps without collapses or oscillations

🎯 Performance: State-of-the-art results at 1.5B scale, matching or exceeding more complex approaches

💰 Efficiency: Comparable or better performance with 2× less compute than multi-stage methods

🔓 Open: Complete evaluation scripts, and model weights released

JustRL/
├── evals/                   # Evaluation scripts
│   ├── gen_vllm.py          # Generation script using vLLM
│   ├── grade.py             # Grading script with hybrid verification
│   └── utils.py             # Answer verification utilities
├── data/                    # Benchmark datasets
│   ├── AIME24/
│   ├── AIME25/
│   ├── AMC23/
│   ├── MATH-500/
│   ├── Minerva/
│   ├── Olympiad-Bench/
│   ├── BRUMO25/
│   ├── CMIMC25/
│   └── HMMT25/
└── justrl_eval_outputs/      # Evaluation outputs (download from Google Drive)
    ├── JustRL-DeepSeek-1.5B/
    │   ├── *.jsonl           # Generation outputs per benchmark
    │   └── grading_results.json
    └── JustRL-Nemotron-1.5B/
        ├── *.jsonl
        └── grading_results.json

We recommend using a conda environment with the following key dependencies:

conda create -n justrl python=3.10
conda activate justrl

• PyTorch: 2.6.0
• vLLM: 0.8.4
• transformers: 4.51.3
• sympy: 1.13.1
• pylatexenc: 2.10

The evaluation outputs are large and hosted on Google Drive. Download them for reproduction:

📥 Download Link: Google Drive

After downloading, extract the justrl_eval_outputs/ directory to the repository root directory.

This evaluation script is based on POLARIS, with one key modification: we add a model-based verifier (CompassVerifier-3B) for more robust evaluation, complementing the rule-based verification system.

cd evals
python gen_vllm.py

Configure the model name in gen_vllm.py by setting the NAME variable. And set appropriateavailable_workers.

cd evals
python grade.py

The grading script processes all JSONL files in the output directory and generates grading_results.json.

Besides, the real question is whether our simplicity comes at a computational cost. It doesn't. We match half of ProRL-V2's compute budget while using a single-stage recipe with fixed hyperparameters. BroRL requires 4.9× more compute by increasing rollouts to 512 per example, essentially exhaustively exploring the solution space. Our approach achieves competitive performance without this computational overhead.

We achieve 64.32% average, slightly outperforming QuestA's 63.81% and leading on five of nine benchmarks. The gap is narrow, which makes sense—both approaches are pushing the boundaries of what's achievable at 1.5B scale. The key difference is in how we get there. We use 2× less compute while achieving slightly better average performance without designing a complex curriculum as used in QuestA.

Our approach is deliberately minimal:

Core Algorithm: Standard GRPO with binary outcome rewards

• Reward: Simple DAPO verifier (string-matching, no SymPy)
• Training: Single-stage, no curriculum or stage transitions
• Hyperparameters: Fixed throughout (no adaptive schedules)
• Data: DAPO-Math-17k without filtering or dynamic sampling
• Length Control: 16K context cap (no explicit penalties)
• Stabilization: Only "clip higher" for gradient stability

Detail hyperparameters and comparisons on training techniques with other methods can refer to our blog.

Training Data: We train on DAPO-Math-17k, a curated dataset of mathematical problems. No offline difficulty filtering or online dynamic sampling is used.

@misc{he2025justrlscaling15bllm,
      title={JustRL: Scaling a 1.5B LLM with a Simple RL Recipe}, 
      author={Bingxiang He and Zekai Qu and Zeyuan Liu and Yinghao Chen and Yuxin Zuo and Cheng Qian and Kaiyan Zhang and Weize Chen and Chaojun Xiao and Ganqu Cui and Ning Ding and Zhiyuan Liu},
      year={2025},
      eprint={2512.16649},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2512.16649}, 
}