Source-linked AI summary
Sparking Scientific Creativity via LLM-Driven Interdisciplinary Inspiration
Priyanka Kargupta, Shuhaib Mehri, Dilek Hakkani-Tur, Jiawei Han
TL;DR
Interdisciplinary research has high impact potential, but existing AI discovery systems often bypass the exploratory reasoning needed for creative, grounded cross-domain ideation. Idea-Catalyst structures this process by decomposing research goals, identifying unresolved conceptual challenges, retrieving and recontextualizing source-domain insights, and prioritizing interdisciplinary ideas; evaluations report higher novelty and insightfulness while retaining problem grounding.
Problem
Interdisciplinary research can have greater long-term impact, yet research remains siloed and existing AI approaches often favor execution or solution design over exploratory, collaborative ideation.
Method
Idea-Catalyst decomposes target-domain goals into research questions, identifies unresolved conceptual challenges, retrieves analogous source-domain insights, and recontextualizes them into incomplete idea fragments.
Results
Idea-Catalyst produces 21.38% more novel and 16.22% more insightful ideas than baseline approaches while remaining grounded in the target research problem.
Takeaways & Limitations
Supporting structured interdisciplinary ideation is presented as a promising direction for human–AI collaboration in scientific research.
Takeaways & Limitations
The framework restricts source domains to sufficiently distant fields and excludes closely related subfields; creative-ideation evaluation is subjective and requires substantial domain expertise.
Abstract
from arXiv · showhide
Despite interdisciplinary research leading to larger and longer-term impact, most work remains confined to single-domain academic silos. Recent AI-based approaches to scientific discovery show promise for interdisciplinary research, but many prioritize rapidly designing experiments and solutions, bypassing the exploratory, collaborative reasoning processes that drive creative interdisciplinary breakthroughs. As a result, prior efforts largely prioritize automating scientific discovery rather than augmenting the reasoning processes that underlie scientific disruption. We present Idea-Catalyst, a novel framework that systematically identifies interdisciplinary insights to support creative reasoning in both humans and large language models. Starting from an abstract research goal, Idea-Catalyst is designed to assist the brainstorming stage, explicitly avoiding premature anchoring on specific solutions. The framework embodies key metacognitive features of interdisciplinary reasoning: (a) defining and assessing research goals, (b) awareness of a domain's opportunities and unresolved challenges, and (c) strategic exploration of interdisciplinary ideas based on impact potential. Concretely, Idea-Catalyst decomposes an abstract goal (e.g., improving human-AI collaboration) into core target-domain research questions that guide the analysis of progress and open challenges within that domain. These challenges are reformulated as domain-agnostic conceptual problems, enabling retrieval from external disciplines (e.g., Psychology, Sociology) that address analogous issues. By synthesizing and recontextualizing insights from these domains back into the target domain, Idea-Catalyst ranks source domains by their interdisciplinary potential. Empirically, this targeted integration improves average novelty by 21% and insightfulness by 16%, while remaining grounded in the original research problem.
1 Introduction
Interdisciplinary creativity develops through iterative synthesis, yet research remains largely siloed and existing AI systems often prioritize execution over exploratory reasoning. Idea-Catalyst addresses this gap by structuring target-domain analysis, cross-domain exploration, and strategic prioritization for creative ideation.
- Motivation: Scientific breakthroughs emerge through gradual synthesis of fragmentary ideas that support discussion, critique, collaboration, and more mature research directions.Early conceptual fragments often originate from multiple domains and help expose conceptual gaps and unresolved challenges.
- Motivation: Interdisciplinary research can produce greater long-term impact, but deeply integrative collaboration remains uncommon across distant fields.Each additional discipline is associated with approximately 20% higher citation impact, while only 5% of cross-domain work involves high-involvement collaboration across non-neighboring fields.
- Prior limitations: Human ideas are grounded in literature and constraints but often remain within familiar single-domain formulations, whereas LLM ideas are cross-disciplinary yet can be superficial or poorly grounded.The contrast creates a complementary tension between grounding and interdisciplinary novelty.
- Prior limitations: Execution-coupled ideation improves feasibility but can encourage incremental, single-domain convergence, while execution-plan systems bypass exploratory and collaborative ideation.These trade-offs motivate augmenting early creative reasoning rather than anchoring immediately on end-to-end automated solutions.
- Idea-Catalyst: Idea-Catalyst decomposes a research problem into core questions, analyzes target-domain progress and unresolved challenges, retrieves analogous insights from source domains, and recontextualizes them into ranked idea fragments.The framework distinguishes domain-specific challenges from deeper conceptual challenges and prioritizes novelty alongside relevance to the original research goal.
- Contributions: Idea-Catalyst produces 21.38% more novel and 16.22% more insightful ideas while remaining grounded in the target research problem.The contribution is supported by LLM-based and human evaluations.
2 Related Work
Prior work links interdisciplinary synthesis to scientific impact and explores automated and human-centered LLM support for research ideation. These approaches motivate structured target-domain analysis and guided cross-domain retrieval to make interdisciplinary inspiration more meaningful.
- Interdisciplinary Research and Scientific Creativity: Interdisciplinary research is associated with scientific innovation and long-term impact, but integrative work across distant fields remains rare and difficult for individual researchers.A central practical challenge is identifying relevant external domains and translating their ideas into the target domain.
- Automated Scientific Discovery and Research Ideation: Automated scientific-discovery systems support literature review, hypothesis generation, research ideation, and experimental planning, but LLM-generated ideas can lack technical depth and feasibility.The passage characterizes these systems as literature-grounded while identifying limitations in their generated ideas.
- Idea-Catalyst: Idea-Catalyst is illustrated as a four-stage framework that analyzes target-domain progress, identifies unresolved challenges, explores analogous source-domain insights, and integrates them into idea fragments.The figure is presented through a real case study summarized in the caption.
- Human-Centered Scientific Knowledge Discovery: Human-centered systems use LLMs for literature exploration, question formulation, iterative refinement, and interdisciplinary information seeking.Their interactive, human-in-the-loop design supports exploratory thinking, while the passage indicates reliance on user input or the LLM’s partial capabilities.
3 Methodology
Idea-Catalyst augments early-stage interdisciplinary ideation by analyzing unresolved target-domain challenges, retrieving conceptually analogous insights from source domains, and recontextualizing them into incomplete idea fragments.
- Preliminaries: Idea-Catalyst defines target and source domains separately, using external fields as reservoirs of transferable, literature-grounded conceptual insights.Source domains are restricted to sufficiently distant fields to promote non-trivial interdisciplinary connections.
- Metacognitive reasoning: The framework grounds creative reasoning in systematic target-domain analysis while preserving exploratory reasoning that expands the solution space.This balances critical evaluation of what is known with creative synthesis of novel and valuable ideas.
- Critical reasoning over the target domain: The framework decomposes a short research problem into target-domain questions and analyzes literature to identify uneven progress and unresolved challenges.Each question has domain-specific and domain-agnostic forms, supporting both precise target-domain assessment and cross-domain comparison.
- Creative reasoning across source domains: Domain-agnostic formulations abstract away terminology and implementation details so external-domain searches can target underlying conceptual gaps.This emphasizes conceptual relevance rather than methodological similarity when selecting source domains.
- Target–source interdisciplinary integration: An idea fragment links a target-domain challenge, source-domain takeaways and literature, and a rationale for how the takeaway could address the challenge.Fragments remain intentionally incomplete so they support creative ideation rather than prescribe full solutions.
4 Experimental Design
The evaluation uses CHIMERA examples, structured preprocessing, retrieval-based baselines, ablations, and LLM preference judgments to assess interdisciplinary ideation quality at takeaway and idea levels.
- Dataset and setup: Idea-Catalyst is evaluated on 400 cross-domain CHIMERA instances linking target-domain contributions with annotated source-domain inspirations.The dataset is designed to study cross-domain knowledge transfer and interdisciplinary ideation.
- Baselines: The comparison includes Free-Form Source Retrieval and Guided Dual-Retrieval, representing increasingly structured LLM ideation approaches.The baselines differ in whether they explicitly analyze the target domain and decompose the research problem.
- Ablations: The study isolates component contributions by removing decomposition, replacing interdisciplinary-potential ranking, and adding conceptual rewriting.These ablations test decomposition, ranking, and output-clarity effects separately.
- Evaluation metrics: Evaluation uses LLM preference judgments against ground truth across takeaway insightfulness and relevance, plus idea novelty and usefulness.Overall win rates indicate which output better satisfies each evaluation criterion.
5 Experimental Results
Idea-Catalyst consistently improves exploratory ideation, producing more novel and insightful ideas while revealing trade-offs between creativity, relevance, and usefulness. Human researchers also found it useful for surfacing research questions and cross-domain insights, though verbosity remained a challenge.
- Overall Performance: Idea-Catalyst achieves the highest insightfulness across top-k settings, improving average relative insightfulness by 16.22% over Guided Dual and 282.21% over Free-Form Source.At the idea level, it also improves average relative novelty by 21.38% over Guided Dual and 407.65% over Free-Form Source.
- Ablation Studies: Removing target-domain decomposition reduces novelty and insightfulness, while replacing interdisciplinary-potential ranking with relevance heuristics lowers overall performance.These ablations support the roles of structured challenge analysis and pairwise interdisciplinary selection.
- Source Domain Distribution: Idea-Catalyst explores diverse source domains while selecting insights for conceptual value rather than diversity alone.It spans Psychology, Biology, Physics, Linguistics, Engineering, and other fields, with normalized entropy H_norm = 0.682.
- Qualitative Comparison: Idea-Catalyst produces more targeted interdisciplinary takeaways than Guided Dual for human–AI collaboration, addressing reciprocal information flow, role adaptation, and learning dynamics.Both methods identify Psychology as the most relevant source domain, but Guided Dual’s Theory of Mind takeaway is more generic.
- Human Study: In a study of six PhD researchers, Idea-Catalyst was viewed as a useful ideation aid, especially for identifying meaningful research questions and surfacing interdisciplinary insights.Retrieved papers received a 3.50/5 rating, while participants rated source-domain takeaways 3.13/5 for relevance and 3.16/5 for insightfulness.
6 Conclusion
The paper introduces Idea-Catalyst as a metacognition-driven framework for targeted interdisciplinary research ideation. Results indicate that it produces more novel and insightful, problem-grounded ideas and may support human–AI collaboration in scientific research.
- Conclusion: Idea-Catalyst guides targeted creative exploration by analyzing target-domain challenges and strategically directing inspiration across scientific domains.The framework is designed to support interdisciplinary research ideation rather than only automated solution generation.
- Conclusion: The framework surfaces ideas and insights that are significantly more novel and insightful than existing baselines while remaining grounded in the original research problem.The paper identifies personalized summarization and interdisciplinary collaborator recommendation as future directions.
A Idea Fragment Output Format
Idea fragments are represented with structured fields covering their title, insight, integration mechanism, selected takeaways, synthesis, and challenge resolution. The format records how source and target concepts combine and contribute to the research problem.
- Idea Fragment Schema: Each idea fragment includes a brief title and a 2–3 sentence core insight describing the integration.The title is specified as brief and descriptive, with a maximum of 15 words.
- Integration Mechanism: The integration mechanism records target-domain elements, source-domain formulations, mechanism explanations, and selection rationales.These fields explain which concepts are combined, how the source framing is used, and why the takeaway is relevant.
- Synthesis: The synthesis approach describes how the selected elements are combined into an integrated idea.This field connects the component concepts into the fragment’s overall conceptual logic.
- Challenge Resolution: Challenge-resolution fields specify how the integration addresses target challenges, source limitations, and the overall research problem.The schema also records concrete realization and key innovations enabled by the integration.
B Baselines
The evaluation compares Idea-Catalyst with baselines that provide progressively more retrieval structure. The baselines differ in whether they analyze target-domain challenges, decompose the problem, and strategically guide source-domain selection.
- Free-Form Source Retrieval: Free-Form Source Retrieval directly identifies source domains, retrieves papers, and synthesizes ideas without explicit target-domain analysis or problem decomposition.It imposes no restriction on source-domain distance from the target domain.
- Guided Dual-Retrieval: Guided Dual-Retrieval first retrieves target-domain literature, then conditions cross-domain exploration and ideation on that context.It remains a retrieve-then-ideate pipeline without explicit unresolved-challenge identification, domain-agnostic abstraction, or strategic source selection.
- Ablations: The ablations remove key Idea-Catalyst components to isolate their contributions.The No Decomposition condition removes structured research-question decomposition and target-domain retrieval conditioned on those questions.
C Domains Supported by Semantic Scholar for Retrieval
Idea-Catalyst retrieves literature through Semantic Scholar’s fixed set of coarse-grained scientific domains. Fine-grained input subfields are mapped to these broader domains for retrieval and filtering.
- Semantic Scholar’s retrieval API searches papers across a fixed set of coarse-grained scientific domains.
- Fine-grained target-domain subfields are mapped to corresponding coarse-grained domains before retrieval.The mapping supports retrieval compatibility with Semantic Scholar’s domain constraints.
- The domain mapping affects literature retrieval and filtering but not conceptual research-question formulation or domain-agnostic abstractions.
D Human Study Details & Participant Backgrounds
The human study involved six PhD researchers from Machine Learning, Natural Language Processing, and Electrical Engineering, each contributing a real research problem from their own work.
- Six PhD researchers in Machine Learning, Natural Language Processing, and Electrical Engineering participated in the human study.
- Each participant provided a real research problem drawn from their own work.
- The study was declared exempt after Institutional Review Board review.
E Evaluation Prompts
The evaluation prompts compare methods and ideas on interdisciplinary insightfulness, relevance, novelty, and usefulness, while emphasizing groundedness, scope appropriateness, and potential impact over polish or immediate practicality.
- Takeaway Evaluation: Takeaways are evaluated for insightful perspectives that introduce specific, intellectually thought-provoking concepts from source domains distinct from the target domain.
- Takeaway Evaluation: Interdisciplinary relevance is judged by potential integration into the target domain and ability to inspire approaches or address research challenges.
- Evaluation Interface: The evaluation interface includes separate screenshots for reviewing questions or challenges, takeaways, and ideas.
- Takeaway Evaluation: Evaluation considers consistency, groundedness, and scope appropriateness, while ignoring explanation length, narrative polish, and missing implementation details.
- Research-Idea Evaluation: Research ideas are compared for interdisciplinary novelty based on source-domain distance, non-obvious approaches, key innovations, and credible surprise.
- Research-Idea Evaluation: Interdisciplinary usefulness depends on integrating source and target concepts to address significant problems or gaps, rather than on simplicity, practicality, or direct applicability.