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.
Figures & tables
Figure 1. Interdisciplinary process of formalizing RL ( Sutton et al., 1998 ) .
Figure 2. Idea-Catalyst is a metacognition-driven framework that (a) analyzes target-domain progress, (b) identifies unresolved challenges, (c) explores source domains for analogous insights, and (d) integrates them into interdisciplinary idea fragments. The figure is illustrated with a real case study (summarized).
Domain-Specific Question ( qiD )
Domain-Agnostic Question ( q1′ )
How can models be trained to dynamically infer and adapt to user intent and task context in real-time collaborative scenarios?
How can understanding of intent and context be updated through continuous interaction?
How should a system decide when to take initiative vs. defer to human to maintain well-calibrated autonomy & control across contexts?
When should control be exercised versus withheld?
Table 1. Output examples of domain-specific & agnostic research question pairs for “human-AI collaboration”.
Domain-Specific Challenge ( q11,D )
Domain-Agnostic Challenge ( q1′1 )
How can a system adapt in real-time to high inter/intra-user variability?
How can behavior adapt to diverse collaborators & evolving goals/environments?
Table 2. Outputted dual representation of a remaining challenge qji in addressing qi , derived from analysis of Dtarget .
Source Domain Concept
How Does it Work?
Metacontrol State Model : Goal-directed behavior reflects a balance between persistence (maintaining current goals) and flexibility (switching goals when conditions change).
Dynamic regulation between persistence and flexibility allows adaptive behavior that remains focused while responding efficiently to changing goals and environments ( Castro et al., 2021 ) .
Cognitive Control : Control processes should be prospectively adjusted based on the expected frequency of goal switches.
Anticipatory adjustment of cognitive control reduces the cost of goal switching, enabling smoother transitions and lower cognitive load under frequent change ( Grahek et al., 2023 ) .
Dynamic Goal Pursuit : Goal pursuit can be supported through just-in-time adaptive interventions that monitor behavior and provide context-sensitive support.
Real-time monitoring and adaptive feedback help sustain goal pursuit under variability by aligning support with evolving goals and situational demands ( O’Driscoll et al., 2024 ) .
Table 3. Summarized conceptual takeaways t1,t2∈Tpsychology for challenge q′11 in Table 2 and their top identified paper.
Method
Insightfulness
Relevance
Overall
@1
@2
@3
@1
@2
@3
@1
@2
@3
Free-Form Source
18.25
23.06
27.35
47.75
51.45
51.57
44.25
48.06
50.14
Guided Dual
73.00
74.19
72.36
62.25
59.68
60.54
66.75
63.23
64.39
Idea-Catalyst
85.50
85.16
84.47 †
60.25
62.42
61.25
63.75
66.45 †
65.67
× Decompose
84.00 †
84.88 †
83.80
60.75 †
61.63
65.32
65.25 †
66.82
70.13
× Potential Ranking
84.71
84.28
84.83
59.40
59.97
61.13
64.16
64.18
65.03
Table 4. Takeaway-level average win rates at top- k ( k∈{1,2,3} ). Bold indicates best; † denotes second-highest.
Method
Novelty
Usefulness
Overall
@1
@2
@3
@1
@2
@3
@1
@2
@3
Free-Form Source
13.50
17.26
19.80
35.50
39.84
40.31
30.50
34.52
35.19
Guided Dual
68.00
70.32
67.95
70.25
65.48 †
64.39
72.25
68.39
66.10
Idea-Catalyst
83.25
84.03
83.05
65.75 †
66.13
66.38
71.25 †
70.65
70.09
× Decompose
82.00 †
81.94
78.48
61.00
63.43
65.06 †
63.75
65.91
67.85 †
× Potential Ranking
83.21 †
83.31 †
81.79 †
62.41
65.15 †
64.16
68.17
69.04 †
67.92
Table 5. Idea-level average win rates at top- k ( k∈{1,2,3} ). Bold indicates best; † denotes second-highest.
Method
Source Domain Framing
Applicability to Target Domain
Guided Dual
Theory of Mind (ToM) and its role in predicting others’ mental states
By equipping AI with ToM capabilities, the system can better understand and predict user intentions, leading to more natural and effective collaboration. This reduces cognitive load by aligning AI behaviors with user expectations and improving task efficiency.
Idea-Catalyst
Reciprocal information flow and role distribution enhance joint action coordination by allowing individuals to dynamically assign and shift roles based on task demands and the predictability of others’ actions.
When individuals engage in reciprocal information flow, they can dynamically assign roles (e.g., leader–follower) and adjust their strategies in real time based on the actions and predictability of their partner. This enables them to adapt to complex and undefined tasks without predefined boundaries.
Table 6. Qualitative Comparison of Source-Domain Takeaways. Exact outputs taken from Guided Dual and Idea-Catalyst .
Figure 3. Source-domain distributions (log-scale) for each method’s top three ideas.
Figure 4. Target-source flow of interdisciplinary inspiration.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Primary Research Area
Target Domain
Research Problem Description
Natural Language Processing
Multilingual NLP
Are there tasks where LLM performance varies systematically across languages, particularly for culture-specific queries, and how can such performance disparities be mitigated?
Electrical Engineering
In-Memory Computing
How can the accuracy of in-memory computing systems be improved while preserving high energy efficiency, especially for Edge-AI applications?
Natural Language Processing
Multilingual Semantics
Why do language models change their answers across languages for the same query, even in high-resource settings, and does this reflect knowledge or semantic misalignment?
Natural Language Processing
Persuasion and Safety
How can we characterize and mitigate the susceptibility of LLMs to persuasion, including harmful or adversarial influence, while preserving beneficial adaptability?
Machine Learning
Model Interpretability
How can influence functions be made dynamic, such that the importance of data points adapts across models and training contexts rather than remaining static?
Natural Language Processing
User Simulation
How can LLM-based user simulators better maintain consistent personas, reflect diverse user behaviors, and be evaluated for realism in multi-turn interactions?
Appendix
Table 7. Participant backgrounds and research problems used in the human study. All participants are PhD researchers ranging from 3 to 5 years of research experience.
Figure 5. Screenshot of evaluation interface (reviewing questions/challenges).
Figure 6. Screenshot of evaluation interface (reviewing takeaways).
Figure 7. Screenshot of evaluation interface (reviewing ideas).
Research idea generation uses computational methods, especially language models, to propose and assess new questions, hypotheses, and research directions.
Large language models (LLMs) excel at structured, verifiable tasks, but their low-entropy bias can produce homogeneous and predictable outputs, limiting their utility for open-ended scientific ideation. Effective discovery, however, spans a broader creative spectrum: from structured day science to loosely structured, serendipitous night science that reaches ideas beyond those typically considered. We introduce AI Night-Scientist, an agentic framework that uses reinforcement learning to teach models when and how to depart from predictable reasoning. Grounded in cognitive science, we model creativity along three axes: action (what to do and how creatively), process (when to explore versus exploit), and outcome (the novelty and usefulness of the resulting idea). We use these axes to train models with GRPO, exposing them to varying degrees and forms of creativity throughout training. This produces substantially more diverse scientific proposals, expanding the range of research directions by 27.8% and contribution types by 14.9% over the base model. It also improves predicted citation impact by up to 32.0 percentage points and originality by 66.2 points. These gains cannot be reproduced by simply increasing decoding temperature; instead, we find that semantic guidance specifying what kind of creativity to pursue is critical. Overall, our results suggest that creativity is a learnable, multi-level ability that can be shaped to help researchers reach ideas beyond those typically explored by LLMs.
Despite rapid progress in automating scientific research, generating promising and well grounded research solutions remains a central challenge. We isolate research ideation as a standalone task and build our solution on the intuition that a challenge in one field can often be addressed by a mechanism that solved an analogous challenge in another. Accordingly, we introduce IdeaScientist, which decomposes ideation into gap finding, innovation, and report writing, and trains each role with reinforcement learning. These roles identify limitations in related work, draw solution intuitions from analogous problem settings, and develop those intuitions into complete research proposals. To facilitate discovery of insights across domains, we construct the Svalbard Idea Vault, a corpus of 2.77M decomposed research ideas for retrieval, training, and temporally controlled evaluation. Our evaluation restricts access to literature available before a cutoff date and assesses how closely proposed directions align with those later explored in 15K papers authored by human researchers. On Qwen3.6-27B, IdeaScientist outperforms the strongest open-source autoresearch baseline by 14.0%, driven mainly by gains in novelty. On this 27B open backbone, IdeaScientist even outperforms Claude Code SDK with Claude-4.8-Opus and Codex SDK with GPT-5.4, by up to 5.9%.
Jiarui Liu, Renjie Tao, Yiwei Liao +15
Carnegie Mellon University · Meta Reality Labs · FAIR at Meta +5
Scientific discovery is an extended process of ideation -- surveying prior work, forming hypotheses, and refining reasoning -- yet existing LLM research agents treat this phase as a brief preamble despite its central role in research. We introduce SciSense, a sensemaking-grounded framework that operationalizes ideation as a structured sequence of eight cognitive stages (Pirolli and Card, 2005). We construct SciSense-Traj, a dataset of 120K citation-conditioned research trajectories in two modes: Hindsight, where an LLM reconstructs the ideation path from a paper's cited works to the paper itself, and Foresight, where the LLM proposes novel directions from the same citations. We distill these into SciSense-LM, a family of sensemaking LLMs spanning 3B to 70B parameters. Contrary to the assumption that looser supervision promotes exploration, Hindsight-trained models improve trajectory quality by 2.0% over Foresight-trained models while also producing more novel and diverse outputs, across all eight model families we test. The advantage propagates downstream: coding agents conditioned on Hindsight trajectories produce research artifacts with higher executability, scientific grounding, and utility than those conditioned on Foresight trajectories. This suggests that outcome-anchored ideation reduces the cognitive burden on downstream agents, freeing them to explore more creatively. SciSense offers both a practical tool for augmenting LLM-driven research workflows and a principled testbed for studying how planning shapes scientific discovery.