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.
Both models use identical hyperparameters without per-model tuning, demonstrating the robustness of our approach.
The AIME24 performance curve for scaling from a weak base DeekSeek-R1-Distill-Qwen-1.5B and a strong base OpenMath-Nemotron-1.5B over thousands of steps.
Key Highlights
✨ 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
Performance
JustRL-DeepSeek-1.5B (Based on DeepSeek-R1-Distill-Qwen-1.5B)
Model
AIME24 (@32)
AIME25 (@32)
AMC23 (@32)
MATH-500 (@4)
Minerva (@4)
OlympiadBench (@4)
HMMT25 (@32)
BRUMO25 (@32)
CMIMC25 (@32)
Avg
DeepSeek-R1-Distill-1.5B
29.90
22.40
63.82
84.90
34.65
45.95
13.44
30.94
12.89
37.65
DeepScaleR-1.5B-Preview
40.21
28.65
73.83
89.30
39.34
52.79
18.96
40.00
21.00
44.88
ProRL-V2
51.87
35.73
88.75
92.00
49.03
67.84
19.38
47.29
25.86
53.08
BroRL
57.50
36.88
/
92.14
49.08
61.54
/
/
/
/
JustRL-DeepSeek-1.5B
52.60
38.75
91.02
91.65
51.47
67.99
21.98
52.71
25.63
54.87
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.
JustRL-Nemotron-1.5B (Based on OpenMath-Nemotron-1.5B)
Model
AIME24 (@32)
AIME25 (@32)
AMC23 (@32)
MATH-500 (@4)
Minerva (@4)
OlympiadBench (@4)
HMMT25 (@32)
BRUMO25 (@32)
CMIMC25 (@32)
Avg
OpenMath-Nemotron-1.5B
58.75
48.44
90.55
92.40
26.93
71.70
30.10
61.67
30.08
56.74
QUESTA-Nemotron-1.5B
71.56
62.08
93.44
92.95
32.08
72.28
40.94
67.50
41.48
63.81
JustRL-Nemotron-1.5B
69.69
62.92
96.02
94.15
30.24
76.59
40.63
66.88
41.72
64.32
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.
Training Recipe
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 paper.
Training Data
We train on DAPO-Math-17k, a curated dataset of mathematical problems. No offline difficulty filtering or online dynamic sampling is used.
Usage
Basic Inference
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
23model_name ="hbx/JustRL-Nemotron-1.5B"# or JustRL-DeepSeek-1.5B4model = AutoModelForCausalLM.from_pretrained(5 model_name,6 torch_dtype="auto",7 device_map="auto"8)9tokenizer = AutoTokenizer.from_pretrained(model_name)1011prompt ="""<problem>
1213Please reason step by step, and put your final answer within \\boxed{}."""1415messages =[{"role":"user","content": prompt}]16text = tokenizer.apply_chat_template(17 messages,18 tokenize=False,19 add_generation_prompt=True20)2122inputs = tokenizer([text], return_tensors="pt").to(model.device)23outputs = model.generate(24**inputs,25 max_new_tokens=16384,26 temperature=0.7,27 top_p=0.9,28 do_sample=True29)3031response = tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True)32print(response)
Batch Inference with vLLM
python
1from vllm import LLM, SamplingParams
23llm = LLM(4 model="hbx/JustRL-Nemotron-1.5B",5 tensor_parallel_size=1,6 max_model_len=327687)89sampling_params = SamplingParams(10 temperature=0.7,11 top_p=0.9,12 max_tokens=16384,13)1415problems =[...]# Your list of problems16responses = llm.generate(problems, sampling_params)
Reproduction
We provide evaluation scripts based on POLARIS, the evaluation script is here
bibtex
1@article{he2025justrl,
2 title={JustRL: Scaling a 1.5 B LLM with a Simple RL Recipe},
3 author={He, Bingxiang and Qu, Zekai and Liu, Zeyuan and Chen, Yinghao and Zuo, Yuxin and Qian, Cheng and Zhang, Kaiyan and Chen, Weize and Xiao, Chaojun and Cui, Ganqu and others},
4 journal={arXiv preprint arXiv:2512.16649},
5 year={2025}
6}
