The application of Large Language Models (LLMs) in accelerating scientific discovery has garnered increasing attention, with a key focus on constructing research agents endowed with innovative capability, i.e., the ability to autonomously generate novel and significant research ideas. Existing approaches predominantly rely on sophisticated prompt engineering and lack a systematic training paradigm. To address this, we propose DeepInnovator, a training framework designed to trigger the innovative capability of LLMs. Our approach comprises two core components. (1) Standing on the shoulders of giants''. We construct an automated data extraction pipeline to extract and organize structured research knowledge from a vast corpus of unlabeled scientific literature. (2) Conjectures and refutations''. We introduce a ``Next Idea Prediction'' training paradigm, which models the generation of research ideas as an iterative process of continuously predicting, evaluating, and refining plausible and novel next idea. Both automatic and expert evaluations demonstrate that our DeepInnovator-14B significantly outperforms untrained baselines, achieving win rates of 80.53%-93.81%, and attains performance comparable to that of current leading LLMs. This work provides a scalable training pathway toward building research agents with genuine, originative innovative capability, and will open-source the dataset to foster community advancement. Source code and data are available at: https://github.com/HKUDS/DeepInnovator.
Figures & tables
Figure 1: The DeepInnovator Framework. Top: We construct training data from arXiv through a carefully designed automated data extraction and synthesis pipeline (Sec. 4.1 ). Bottom: We perform RL training via a meticulously designed next-idea prediction task ( 4.2 ), coupled with a decoupled reward and critique mechanism (Sec. 4.3 ).
Figure 2: Evaluation results of ideas generated by eight models across six rubrics. DeepInnovator outperforms Qwen-14B-Instruct on all dimensions and demonstrates competitive capability against leading LLMs.
Metric
Qwen-14B-Instruct
Deepseek-r1
Gemini-2.5-pro
GPT-4o
Minimax-M2.1
Qwen3-max
Grok-4.1
Novelty
80.53%
47.79%
15.93%
49.56%
59.29%
32.74%
37.17%
Feasibility
87.61%
23.01%
18.58%
11.50%
10.62%
19.47%
10.62%
Effectiveness
92.92%
45.13%
43.36%
76.11%
50.44%
66.37%
43.36%
Detailedness
93.81%
59.29%
68.14%
44.25%
61.06%
91.15%
45.13%
Table 1: Win rate results of DeepInnovator vs Other Models. Novelty : whether the approach is innovative; Feasibility : whether it can be practically implemented; Effectiveness : whether the problem can be solved; Detailedness : whether the proposal is complete and specific. Bold indicate that DeepInnovator wins (i.e., win rate > 50%).
Figure 3: Comparison of win rates between ideas generated by DeepInnovator at each step and the initial idea (step 1). Our training method effectively enhances DeepInnovator’s ability to refine research ideas.
Domain
Metric
Much Better
Better
Worse
Much Worse
Both Bad
Avg. Winrate
Law
Novelty
0 / 0
9 / 7
3 / 6
1 / 0
0 / 0
69.2% / 53.8%
Feasibility
1 / 0
5 / 4
3 / 7
1 / 1
3 / 1
60.0% / 33.3%
Effectiveness
2 / 1
5 / 4
2 / 3
1 / 3
3 / 2
70.0% / 45.5%
Detailedness
7 / 1
3 / 4
2 / 5
1 / 0
0 / 3
76.9% / 50.0%
Education
Novelty
3 / 0
9 / 9
2 / 5
1 / 1
0 / 0
80.0% / 60.0%
Feasibility
3 / 0
5 / 0
5 / 7
1 / 3
1 / 5
57.1% / 0.0%
Table 2: Winrate results annotated by human experts. “Much Better” or “Much Worse” indicates that the expert thinks that DeepInnovator is much better/worse than Qwen-14B-Instruct/GPT-4o. The value before the slash “/” denotes DeepInnovator vs Qwen-14B-Instruct, and the value after the slash denotes DeepInnovator vs GPT-4o. The values represent the sum of annotators’ judgments. Bold indicate that DeepInnovator wins (i.e., win rate > 50%).
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Category
Parameter
Value
Dataset Statistics
Training set size (by category)
cs: 433, math: 194, q-fin: 181, stat: 204
Validation set size (by category)
cs: 51, math: 22, q-fin: 18, stat: 20
Training Data
Train batch size
16
Max prompt length
8192 tokens
Appendix
Table A1: Hyperparameters used in DeepInnovator.
Rubric
Description
Detailed, specific solution
Does the part of the plan relevant to satisfying this rubric item include fully specified details on how to implement it? There should be no claims of handling something without actually doing so, no vague terms, ambiguity, or lack of clarity. The description should use simple, easy-to-understand language.
No overlooked flaws or weaknesses
Are there any important overlooked flaws or weaknesses in the part of the plan addressing this rubric item that would invalidate its claimed satisfaction of the criterion?
Well-justified rationale
Is the part of the plan relevant to this rubric item well-motivated and justified? For example, does it provide convincing arguments that the chosen approach is better than simpler alternatives or competing hypotheses?
Cost and effort efficient
Does the plan address this rubric item in a cost- and effort-efficient manner, avoiding unnecessary complexity? Consider whether a less resource-intensive or less labor-demanding solution could achieve comparable effectiveness.
No ethical issues
Does this part of the plan pose any potential for negative societal impact or raise ethical concerns (e.g., bias, privacy violations, manipulation, or unfair outcomes)?
Consistent with overall plan
Is this component of the plan logically consistent with the rest of the proposed strategy? Ensure it does not contradict other stated assumptions, methods, or objectives elsewhere in the plan.
Appendix
Table A2: Evaluation Rubrics.
ID
Reason
Model Generating the Issue
Law1
Incomplete idea generation
Qwen-14B-Instruct
Law2
Incomplete idea generation
Qwen-14B-Instruct
Law3
Incomplete idea generation
Qwen-14B-Instruct
Law4
Generated idea is a hybrid of law and computer science
DeepInnovator
Biotech1
Incomplete idea generation
Qwen-14B-Instruct
Appendix
Table A3: Issue Table
Position
Count
Postdoc
1
PhD
6
Mphil
2
Appendix
Table A4: Positions of the idea judge participants.
Institution
Count
The University of Hong Kong (HKU)
3
Zhejiang University (ZJU)
3
Broad Institute of MIT & Harvard (Broad Institute)
1
Swiss Federal Technology Institute of Lausanne (EPFL)
1
The California Institute of Technology (Caltech)
1
Appendix
Table A5: Institutions of the idea judge participants.
Real Ideas
Fake Ideas
Accuracy (%)
80.77
75.12
Appendix
Table A6: Discrimination Accuracy of Qwen-plus on arXiv ideas (Real) vs. Qwen-14B-Instruct Generated Ideas (Fake)
Scientific progress depends on the continual generation of innovative re-search ideas. However, the rapid growth of scientific literature has greatly increased the cost of knowledge filtering, making it harder for researchers to identify novel directions. Although existing large language model (LLM)-based methods show promise in research idea generation, the ideas they produce are often repetitive and lack depth. To address this issue, this study proposes a multi-agent iterative planning search strategy inspired by com-binatorial innovation theory. The framework combines iterative knowledge search with an LLM-based multi-agent system to generate, evaluate, and re-fine research ideas through repeated interaction, with the goal of improving idea diversity and novelty. Experiments in the natural language processing domain show that the proposed method outperforms state-of-the-art base-lines in both diversity and novelty. Further comparison with ideas derived from top-tier machine learning conference papers indicates that the quality of the generated ideas falls between that of accepted and rejected papers. These results suggest that the proposed framework is a promising approach for supporting high-quality research idea generation. The source code and dataset used in this paper are publicly available on Github repository: https://github.com/ChenShuai00/MAGenIdeas. The demo is available at https://huggingface.co/spaces/cshuai20/MAGenIdeas.
Shuai Chen, Chengzhi Zhang
Department of Information Management, Nanjing University of Science and Technology, Nanjing, 210094 China
Large Language Models (LLMs) have demonstrated potential in automating scientific ideation, yet current approaches relying on iterative prompting or complex multi-agent architectures often suffer from hallucination or computational inefficiency. A critical bottleneck in applying Reinforcement Learning (RL) to this open-ended domain is reward hacking -- where models exploit imperfect evaluation proxies to maximize scores without producing genuine scientific innovation. To address these limitations, we propose an RL framework explicitly tailored for high-quality scientific idea generation. We propose the first multi-agent reward function designed to serve as a judge, decoupling methodological validation from implementation details while providing strict binary rewards that are robust to reward hacking. To effectively optimize against this sparse signal, we utilize an unbiased variant of Group Relative Policy Optimization to mitigate artificial length bias. We grounded our training in ICLR-320, a curated dataset of problem-solution pairs extracted from ICLR 2024 proceedings. Experiments demonstrate that our framework significantly outperforms state-of-the-art baselines across expert-evaluated metrics of novelty, feasibility, and effectiveness.
As automated scientific discovery advances, Large Language Models (LLMs) can now generate research ideas at an unprecedented scale, shifting the bottleneck from idea generation to idea evaluation. Existing evaluators mainly rely on parametric LLM knowledge or unstructured retrieval, producing judgments that lack the experience-grounded reasoning used by human instructors. To address this, we propose DeepInstructor, an agentic framework that formulates idea evaluation as reasoning over structured scholarly experience. DeepInstructor constructs an Experience Graph from 58,607 peer reviews and employs a ReAct-based agent to retrieve dimension-specific evidence for traceable evaluation. We further introduce DeepInstruct, a dataset with controlled pairwise comparisons across novelty, significance, and feasibility. Experiments show that DeepInstructor substantially outperforms existing baselines, improving Hit@1 and Hit@2 alignment with human judgments by 24.4% and 29.7%, respectively. Our findings suggest that scientific idea evaluation can be grounded in explicit reasoning over structured scholarly experience
Rongcan Pei, Fang Guo, Qinglin Qi +7
Tongji University · Westlake University · Zhejiang University +2