MindFlow: Mind Supernet Powered Thinking Flows for Research Idea Innovation

Mengdi Liu; Wenjue Chen; Wenyue Chen; Cheng Yang; Fanqi Kong; Zhangyang Gao; Xiaoxue Cheng; Yiheng Li

cs.AIOct 8, 2026

MindFlow: Mind Supernet Powered Thinking Flows for Research Idea Innovation

Authors: Mengdi Liu, Wenjue Chen, Wenyue Chen, Cheng Yang, Fanqi Kong, Zhangyang Gao, Xiaoxue Cheng, Yiheng Li, +5 more

Organizations: State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, China · University of Chinese Academy of Sciences, China · Peking University · DeepWisdom · Shanghai Artificial Intelligence Laboratory · Renmin University of China

Abstract

Research idea innovation is a fundamental engine of scientific progress, yet it remains difficult to generate and evaluate in a scalable and controllable way. This challenge lies in its inherently open-ended and multi-objective nature, where ideas should balance novelty, plausibility and feasibility. While recent LLM-based approaches have made progress through carefully designed prompts or agent pipelines, they are constrained by predefined, static ideation workflows. To address this limitation, we propose MindFlow, a framework that explicitly formulates ideation as a graph-structured Flow in Mind, which is composed of modular thinking operators and modeled by a probabilistic mind supernet. Given a research topic, a controller dynamically samples thinking flows to generate candidate ideas. This open-ended problem is optimized using a tournament-based relative ranking, enabling the controller to progressively favor higher-quality thinking flows. We further introduce an evaluation protocol that jointly assesses problem finding and problem solving, going beyond title- or abstractonly judgments. Across diverse topics, MindFlow shows its superiority as an explicit, controllable and optimizable research idea innovator.

Figures & tables

Figure 1 : Overview of challenges in research idea generation and comparison between prior pipelines and MindFlow.

Figure 2 : The overall framework of our proposed MindFlow. MindFlow formulates research idea innovation as a graph-structured thinking flow composed of atomic thinking operators and modeled by a probabilistic mind supernet. Given a research topic, a controller dynamically samples a tailored thinking flow to generate a idea, and the controller is optimized using a tournament-based relative ranking.

Method	Problem Finding (Motivation)				Problem Solving (Method)				Overall
Method	Novelty	Significance	Timeliness	MOScore	Novelty	Effectiveness	Feasibility	MOScore	Overall
Generate	0.294	0.602	0.455	0.441	0.071	0.171	0.308	0.169	0.305
GenerateCoT	0.687	0.379	0.469	0.504	0.308	0.076	0.213	0.185	0.344
AI Scientist	0.199	0.768	0.498	0.456	0.028	0.123	0.455	0.159	0.308
AI-Researcher	0.692	0.351	0.450	0.488	0.265	0.081	0.175	0.165	0.326
VIRSCI	0.711	0.645	0.645	0.667	0.445	0.332	0.109	0.274	0.470

Table 1 : Win-rate evaluation under our LLM-judged protocol. We report dimension-wise win rates for problem finding (novelty, significance, and timeliness) and problem solving (novelty, effectiveness, and feasibility), along with their aggregated multi-objective scores (MOScore) and overall average. The best and second-best results are labeled with bold and underline.

Method	Novelty	Diversity	Effectiveness	Feasibility
Generate	0.370	0.184	0.642	0.539
GenerateCoT	0.422	0.275	0.626	0.191
AI Scientist	0.352	0.167	0.663	0.172
AI-Researcher	0.474	0.293	0.589	0.672
VIRSCI	0.457	0.281	0.630	0.583
MindFlow	0.541	0.322	0.665	0.735

Table 2 : Computable evaluation on novelty, diversity, effectiveness, and feasibility, reflecting multi-objective beyond LLM judgments. The best and second-best are labeled with bold and underline.

Method	Problem Finding (Motivation)				Problem Solving (Method)				Overall
Method	Novelty	Significance	Timeliness	MOScore	Novelty	Effectiveness	Feasibility	MOScore	Overall
AIScientist	0.175	0.688	0.425	0.403	0.100	0.213	0.638	0.271	0.337
AI-Researcher	0.600	0.313	0.388	0.428	0.325	0.163	0.275	0.248	0.338
VIRSCI	0.650	0.563	0.575	0.595	0.613	0.450	0.138	0.356	0.475
MindFlow	0.688	0.663	0.650	0.667	0.550	0.538	0.400	0.494	0.580

Table 3 : Human expert evaluation on 50 randomly sampled queries from IdeaBench. We report dimension-wise win rates for problem finding (motivation) and problem solving (method), along with their aggregated multi-objective scores (MOScore) and overall average. The best and second-best results are labeled with bold and underline.

Agreement Type	PF-Novelty	PF-Significance	PF-Timeliness	PS-Novelty	PS-Effectiveness	PS-Feasibility	Average
Inter-judge Agreement Rate	79.3%	68.7%	76.0%	80.7%	72.0%	72.7%	74.9%
Human–LLM Agreement Rate	82.1%	71.4%	78.8%	83.5%	75.3%	75.6%	77.8%

Table 4 : Agreement analysis of the LLM-judged evaluation protocol.

Figure 3 : The visualization of MindFlow’s operator sampling process and visualization of the sampling flow.

Operator	Problem Finding (Motivation)				Problem Solving (Method)				Overall
Operator	Novelty	Significance	Timeliness	MOScore	Novelty	Effectiveness	Feasibility	MOScore	Overall
Divergent + Convergent Thinking	0.569	0.564	0.531	0.554	0.265	0.194	0.180	0.210	0.383
Counterfactual Thinking	0.901	0.412	0.569	0.611	0.569	0.123	0.128	0.240	0.426
Analogical Thinking	0.578	0.716	0.635	0.641	0.271	0.271	0.171	0.234	0.438
Constraint-Driven Thinking	0.791	0.526	0.706	0.669	0.370	0.346	0.142	0.275	0.472
Critical Thinking	0.275	0.796	0.531	0.511	0.062	0.192	0.275	0.162	0.336

Table 5 : Operator-wise win-rate evaluation under our LLM-judged protocol. We report dimension-wise win rates for problem finding (motivation) and problem solving (method), along with their aggregated multi-objective scores (MOScore) and overall average. The best and second-best results are labeled with bold and underline.

Method	CV	NLP	Multimodal	Audio & Speech	Robotics	Science	General ML	Theory
Generate	0.302	0.289	0.334	0.257	0.284	0.327	0.281	0.365
GenerateCoT	0.355	0.366	0.307	0.259	0.419	0.348	0.352	0.372
AI Scientist	0.281	0.318	0.348	0.254	0.344	0.388	0.305	0.315
AI-Researcher	0.348	0.293	0.264	0.317	0.335	0.314	0.342	0.422
VIRSCI	0.502	0.450	0.436	0.324	0.533	0.473	0.468	0.568
MindFlow (Ours)	0.503	0.498	0.480	0.398	0.694	0.412	0.576	0.660

Table 6 : Performance across topic domains ( problem finding and problem solving ). We report per-domain scores under our LLM-judged protocol. The best and second-best results are labeled with bold and underline.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Method	CV	NLP	Multimodal	Audio & Speech	Robotics	Science	General ML	Theory
Divergent +
Convergent Thinking	0.363	0.394	0.374	0.268	0.392	0.359	0.416	0.391
Counterfactual Thinking	0.423	0.400	0.402	0.406	0.490	0.334	0.425	0.539
Analogical Thinking	0.457	0.449	0.434	0.331	0.402	0.409	0.388	0.564
Critical Thinking	0.338	0.350	0.359	0.276	0.345	0.335	0.313	0.433
Constraint-Driven Thinking	0.439	0.473	0.445	0.237	0.628	0.316	0.532	0.601

Appendix

Table 7 : Operator-wise performance across topic domains ( problem finding and problem solving ) under our LLM-judged protocol. We report domain-wise scores for each thinking operator. The best and second-best results are highlighted in bold and underlined, respectively.

	Problem Finding (Motivation)				Problem Solving (Method)				Overall
Method	Novelty	Significance	Timeliness	MOScore	Novelty	Effectiveness	Feasibility	MOScore	Overall
AIScientist	0.175	0.688	0.425	0.403	0.100	0.213	0.638	0.271	0.337
AI-Researcher	0.600	0.313	0.388	0.428	0.325	0.163	0.275	0.248	0.338
VIRSCI	0.650	0.563	0.575	0.595	0.613	0.450	0.138	0.356	0.475
MindFlow	0.688	0.663	0.650	0.667	0.550	0.538	0.400	0.494	0.580

Appendix

Table 8 : Human expert evaluation results.

	Problem Finding (Motivation)			Problem Solving (Method)			Average
	Novelty	Significance	Timeliness	Novelty	Effectiveness	Feasibility	Average
Inter-judge Agreement	79.3%	68.7%	76.0%	80.7%	72.0%	72.7%	74.9%
Human-LLM Agreement	82.1%	71.4%	78.8%	83.5%	75.3%	75.6%	77.8%

Appendix

Table 9 : Agreement analysis across inter-judge agreement among three LLM judges and between human experts and LLM judges .

Method	Novelty	Significance	Timeliness	MOScore
VIRSCI	$0.711\pm 0.018$	$0.645\pm 0.021$	$0.645\pm 0.016$	$0.667\pm 0.015$
MindFlow	$0.744\pm 0.015$	$0.782\pm 0.019$	$0.702\pm 0.017$	$0.742\pm 0.012$

Appendix

Table 10 : Problem Finding (Motivation) performance (mean

\pm

std over three runs).

Method	Novelty	Effectiveness	Feasibility	MOScore
VIRSCI	$0.445\pm 0.023$	$0.332\pm 0.019$	$0.109\pm 0.014$	$0.274\pm 0.017$
MindFlow	$0.361\pm 0.020$	$0.408\pm 0.022$	$0.257\pm 0.018$	$0.339\pm 0.014$

Appendix

Table 11 : Problem Solving (Method) performance (mean

\pm

std over three runs).

Explore similar work

LDC: Learning to Generate Research Idea with Dynamic Control
Dec 19, 2024Ruochen Li, Liqiang Jing, Chi Han +2Controllable Text Generation Research Idea Generation
University of Texas at Dallas · UIUC · Stony Brook University
EvoGens: A Population-Based Heuristic Search Framework for Scientific Idea Generation
May 29, 2026Xu Li, Hanzhe Tu, Xinyi Li +3Scientific Hypothesis Generation Research Idea Generation
Southwest Petroleum University, Chengdu, China · Sichuan Police College, Luzhou, China
Enhancing Research Idea Generation through Combinatorial Innovation and Multi-Agent Iterative Search Strategies
Apr 22, 2026Shuai Chen, Chengzhi ZhangMulti-Agent LLM Systems Computational Creativity
Department of Information Management, Nanjing University of Science and Technology, Nanjing, 210094 China

Dec 19, 2024, cs.CLJ/K move · Enter open · S save

LDC: Learning to Generate Research Idea with Dynamic Control

Ruochen Li, Liqiang Jing, Chi Han, Jiawei Zhou+1

University of Texas at Dallas · UIUC · Stony Brook University

Controllable Text Generation Research Idea Generation Large Language Model-Guided Scientific Discovery

Open paper AskPDF arXiv

Recent advancements in large language models (LLMs) have demonstrated their potential in automating the scientific research ideation. Existing approaches primarily focus on prompting techniques, often producing ideas misaligned with expert standards - novelty, feasibility, and effectiveness, which are widely recognized by the research community as the three key subdimensions of high-quality ideas. Also, balancing these dimensions remains challenging due to their inherent trade-offs. To address these limitations, we propose the first framework that employs a two-stage approach combining Supervised Fine-Tuning (SFT) and controllable Reinforcement Learning (RL) for the task. In the SFT stage, the model learns foundational patterns from pairs of research papers and their corresponding follow-up ideas. In the RL stage, multi-dimensional reward models guided by fine-grained feedback evaluate and optimize the model across key dimensions. During inference, dimensional controllers coordinated by a sentence-level decoder enable dynamic context-aware steering of the idea generation process. Our framework provides a balanced approach to research idea generation, achieving high-quality outcomes in the experiment by dynamically navigating the trade-offs among novelty, feasibility, and effectiveness.

Abstract

Explore similar work

LDC: Learning to Generate Research Idea with Dynamic Control

EvoGens: A Population-Based Heuristic Search Framework for Scientific Idea Generation

Enhancing Research Idea Generation through Combinatorial Innovation and Multi-Agent Iterative Search Strategies