An automated data extraction pipeline that extracts and organizes structured research knowledge from a vast corpus of unlabeled scientific literature.
A "Next Idea Prediction" training paradigm that models the generation of research ideas as an iterative process of continuously predicting, evaluating, and refining plausible and novel next ideas.
1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3model_name = "T1anyu/DeepInnovator"
4tokenizer = AutoTokenizer.from_pretrained(model_name)
5model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
6
7prompt = "Based on the recent advances in graph neural networks and large language models, propose a novel research idea:"
8
9messages = [
10 {"role": "user", "content": prompt}
11]
12
13text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
14inputs = tokenizer([text], return_tensors="pt").to(model.device)
15
16outputs = model.generate(
17 **inputs,
18 max_new_tokens=1024,
19 temperature=0.7,
20 top_p=0.9,
21 do_sample=True,
22)
23
24response = tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True)
25print(response)
1from vllm import LLM, SamplingParams
2
3llm = LLM(model="T1anyu/DeepInnovator")
4sampling_params = SamplingParams(temperature=0.7, top_p=0.9, max_tokens=1024)
5
6prompt = "Based on the recent advances in graph neural networks and large language models, propose a novel research idea:"
7outputs = llm.generate([prompt], sampling_params)
8
9print(outputs[0].outputs[0].text)
Both automatic and expert evaluations demonstrate that DeepInnovator-14B significantly outperforms untrained baselines:
1@article{fan2026deepinnovator,
2 title={DeepInnovator: Triggering the Innovative Capabilities of LLMs},
3 author={Fan, Tianyu and Zhang, Fengji and Zheng, Yuxiang and Chen, Bei and Niu, Xinyao and Huang, Chengen and Lin, Junyang and Huang, Chao},
4 journal={arXiv preprint arXiv:2602.18920},
5 year={2026}
6}
This model is released under the
Apache 2.0 License.
This work is developed by the
HKU Data Science Lab (HKUDS).
