SPARK is a novel reinforcement learning framework that enables autonomous strategic exploration for long-horizon agentic tasks. Instead of uniformly exploring all steps, SPARK selectively branches at critical decision points using intrinsic <explore> signals, achieving superior performance with significantly fewer training samples.

Key Features

🎯 Autonomous Strategic Exploration: Dynamically identifies critical states for branching without human priors
⚡ Sample Efficient: Achieves 84.4% success with only 20% training data (vs. GRPO 76.6% at 100%)
💰 Token Efficient: Reduces token consumption by up to 47% through prefix sharing
🌍 Strong Generalization: Maintains 80.5% success on unseen tasks (significantly outperforms GRPO)

Performance Highlights

Benchmark	SPARK-1.5B	GPT-5	Gemini-2.5-Pro
ALFWorld L2	80.5%	63.3%	55.5%
ScienceWorld L2	49.2%	33.6%	30.5%
WebShop	75.8%	29.7%	32.0%

Quickstart

Here we provide a transformers inference style:

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Jinyang23/Spark-1.5B-ScienceWorld"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Calculate the sum of 123 and 456. Provide only the numerical answer."

messages = [
    {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Citation

If you use this model or the SPARK framework in your research, please cite:

bibtex

@article{wu2026spark,
  title={SPARK: Strategic Policy-Aware Exploration via Dynamic Branching for Long-Horizon Agentic Learning},
  author={Wu, Jinyang and Yang, Shuo and Yang, Changpeng and Shen, Yuhao and Zhang, Shuai and Wen, Zhengqi and Tao, Jianhua},
  journal={arXiv preprint arXiv:2601.20209},
  year={2026}
}

Model Details

Base Model: Qwen/Qwen2.5-1.5B-Instruct
Training Method: SPARK (Dynamic Branching RL)
Training Dataset: ScienceWorld

Jinyang23/Spark-1.5B-ScienceWorld

SPARK: Strategic Policy-Aware Exploration via Dynamic Branching

Overview

Key Features

Performance Highlights

Quickstart

Citation

Model Details

Links

Jinyang23/
Spark-1.5B-ScienceWorld