Idea-Catalyst icon Idea-Catalyst
Boosting creativity in research

Sparking Scientific Creativity via
LLM-Driven Interdisciplinary Inspiration

Priyanka Kargupta, Shuhaib Mehri, Dilek Hakkani-Tur, Jiawei Han University of Illinois Urbana-Champaign
Read Paper Code Dataset

Overview

Most AI-for-science systems optimize for fast solution generation. This paper argues that the more important missing layer is the reasoning process behind creative, interdisciplinary ideation. Idea-Catalyst is built for that earlier stage: helping researchers identify where a field is stuck, where adjacent ideas might transfer, and which cross-domain signals are actually worth exploring.

Start from the bottleneck, not the answer. The key move is to identify an unresolved challenge inside the target field before jumping to ideation.
Borrow the mechanism, not just the concept. Interdisciplinary inspiration becomes useful when an outside domain tackles the same abstract problem.
Recontextualize and prioritize the strongest insights. Extracted source-domain insights are translated back into the target field, turned into candidate idea fragments, and ranked by their potential to complement the target field.
Figure showing reinforcement learning emerging from multiple disciplines
Figure 1. The motivating example is not just that RL touched many fields, but that each field contributed a different piece of the missing reasoning chain.
Behavioral psychology contributes the reward intuition, control theory contributes formal optimization, and animal learning contributes mechanisms for delayed signals. Idea-Catalyst is designed to recreate that style of disciplined cross-domain borrowing.
+21%
Novelty Gain

Average relative improvement over Guided Dual on idea-level novelty.

+16%
Insightfulness Gain

Average relative improvement over Guided Dual on takeaway-level insightfulness.

400
Research Problems

Dataset coverage spanning a broad set of target-domain research goals.

1,730
Interdisciplinary Insights

Extracted cross-domain takeaways used to seed new directions.

1,330
Interdisciplinary Ideas

Recontextualized idea candidates produced across the benchmark.

How Idea-Catalyst Works

The framework is designed around a simple principle: interdisciplinary inspiration should be earned through structured analysis, not guessed through unconstrained retrieval. It operationalizes three metacognitive behaviors from the paper: clarifying research goals, recognizing unresolved challenges, and strategically exploring external domains with high impact potential.

Hover to inspect stage
1

Analyze the target field

Start with a concrete research goal, then unpack it into target-domain questions, current wins, and unresolved challenges.

This is where the system decides what the field still cannot explain or reliably do.
Hover to inspect stage
2

Rewrite the challenge

Translate a domain-specific obstacle into a portable conceptual challenge that other disciplines might already know how to solve.

The abstraction step prevents retrieval from overfitting to surface vocabulary in the target domain.
Hover to inspect stage
3

Rank cross-domain inspiration

Retrieve candidate domains, extract the most relevant takeaways, and rank them by interdisciplinary potential before integrating them back.

The goal is not maximum diversity by itself. The goal is conceptual leverage.
Interactive pipeline walkthrough

Hover a stage above to see what the system is producing, filtering, and carrying forward.

1
Map the target problem
Input goal Effective and reliable human-AI collaboration.
→
Target-domain questions What helps models infer intent? What feedback loops reduce cognitive load?
Known wins vs. remaining gaps The system distinguishes what the field already solves from what remains brittle, underspecified, or non-scalable.
real-time intent inference dynamic user modeling adaptive task prediction
Output of Stage 1: a crisp set of research questions and the specific challenge that still blocks progress.
2
Abstract the bottleneck
Domain-specific challenge Adapting in real time to collaborators with evolving goals and behaviors.
→
Conceptual challenge Dynamic balancing of persistence and flexibility under shifting signals.
Why this abstraction matters Now the search is no longer trapped inside AI wording. Psychology or sociology may have already studied the same underlying problem.
adaptive goal shifting social role adaptation behavior under uncertainty
Output of Stage 2: a searchable challenge description that preserves the bottleneck while dropping field-specific jargon.
3
Select the strongest outside insight
Psychology
0.86
Sociology
0.64
Control Theory
0.72
Top takeaway Psychology suggests balancing persistence and flexibility, while sociology highlights role-aware adaptation to evolving collaborators and interaction settings.
Integrated idea Maintain and update a latent, role-aware, and socially sensitive interaction state, learned through feedback and perspective-taking, so the system can adapt as collaborator goals and behavior shift.
Output of Stage 3: ranked source-domain takeaways that can be reintegrated into a concrete new research direction.

Figure 2: Idea-Catalyst pipeline

Framework diagram for Idea-Catalyst
Figure 2. Showing target-domain analysis, challenge abstraction, source-domain retrieval, ranking, and reintegration into final research ideas.

Results

Across both takeaway-level and idea-level evaluations, the method consistently beats retrieval-only baselines on the dimensions that matter for exploratory research: insightfulness and novelty. More diverse retrieval alone is not enough. What matters is selecting source domains that offer meaningful conceptual leverage for the target problem.

Takeaway Evaluation

Highest insightfulness at every top-k

Idea-Catalyst reaches the strongest takeaway-level insightfulness scores across top-1, top-2, and top-3 comparisons, averaging a 16.22% relative improvement over Guided Dual.

Idea Evaluation

Novelty improves without drifting off-task

On final integrated ideas, the paper reports a 21.38% average relative novelty gain over Guided Dual and substantially stronger performance than free-form source retrieval.

Ablation Signal

Decomposition and ranking both matter

Removing target-domain decomposition or replacing interdisciplinary potential-based ranking weakens performance, suggesting the gains come from structured reasoning rather than cosmetic rewriting.

Source Diversity

Figure 3: Without guidance, ideation stays close to home

Open Figure 3 PDF
Figure 3. This figure compares the distribution of source domains retrieved by Idea-Catalyst and the baseline retrieve-then-ideate methods. Without exploration guidance, LLMs tend to stay within the same domain, or very nearby ones, even when asked for "interdisciplinary" ideas. Idea-Catalyst pushes the search outward, then prioritizes the outside domains that best match the unresolved challenge.
Inspiration Flow

Figure 4: The right source domain depends on the challenge

Open Figure 4 PDF
Figure 4. Different target problems draw on different outside disciplines. The point is not to retrieve from as many domains as possible, but to identify which source domains offer the most meaningful conceptual leverage for the specific target challenge.

Blindly choosing diverse domains is not enough. The strongest interdisciplinary ideas come from choosing the outside domains that best match the unresolved challenge.

Human Study

Positive signal from real researchers

Six PhD researchers evaluated the system on their own problems. Generated research questions received a 4.00/5 average score, retrieved literature scored 3.50/5, source-domain takeaways were judged moderately relevant and insightful, and the final ideas were rated more novel than useful, reflecting the common gap between ideation and execution.

Evaluation Interface

Appendix interface views used in the human study

Interface screenshot showing the research problem setup step
Problem setup and target-question framing.
Interface screenshot showing takeaway evaluation
Takeaway review and source-domain evaluation.
Interface screenshot showing final idea assessment
Final idea assessment with usefulness and novelty judgments.

Citation

If you find our paper to be useful, please cite it! 

@misc{kargupta2026sparkingscientificcreativityllmdriven,
  title         = {Sparking Scientific Creativity via LLM-Driven Interdisciplinary Inspiration},
  author        = {Priyanka Kargupta and Shuhaib Mehri and Dilek Hakkani-Tur and Jiawei Han},
  year          = {2026},
  eprint        = {2603.12226},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2603.12226}
}