PaperlayerView exact revision on arXiv

Source-linked AI summary

Orchestrating GenAI for Interdisciplinary Research

Shirley Anugrah Hayati, Moyan Zhou, Patricia Anugrah Setiani, Ruizi Wang, Joseph Chee Chang, Dongyeop Kang

arXiv:2609.30588v1cs.HC

TL;DR

Prior work leaves the broader practices of GenAI use in interdisciplinary research less understood. In a longitudinal study and interviews with 15 researchers, the authors examined how they orchestrated GenAI across research goals and disciplines. Researchers treated GenAI as an information-foraging assistant while retaining responsibility for evaluation and integration.

  • Problem

    How interdisciplinary researchers use GenAI to navigate and integrate knowledge across fields remains less understood than its use in general or single-discipline research.

    Source: §1 · p. 1, §2.1 · p. 3

  • Method

    The authors studied 15 interdisciplinary researchers using standard GenAI and Deep Research in real-world workflows, triangulating longitudinal records, interaction logs, and interviews.

    Source: GenAI tools to navigate challenges in… · p. 2

  • Results

    Researchers treated GenAI mainly as an information-foraging assistant, shifting between expert, learner, and integrator roles while valuing inspiration and connections as well as correctness.

    Source: §6 · p. 19

  • Takeaways & Limitations

    Researchers still need to evaluate and integrate unfamiliar knowledge into their research context, so research assistants should preserve their epistemic agency.

    Source: §6 · p. 19

    Success ratings captured participants’ perceived helpfulness, not verified correctness of GenAI outputs.

    Source: §5.3 · p. 19

Abstract

from arXiv · show

As researchers tackle interdisciplinary problems, they face the need to deepen expertise in primary areas while rapidly acquiring knowledge in secondary domains. Generative AI (GenAI) is increasingly positioned to meet this need, from general-purpose chat assistants to Deep Research tools marketed as autonomous research agents. Prior work has examined how researchers use GenAI to support single-discipline or general research tasks. However, we know little about the goals and GenAI practices in interdisciplinary research. We conducted a longitudinal study and semi-structured interviews with 15 interdisciplinary researchers to examine how interdisciplinary researchers actually orchestrate GenAI. Findings show that researchers leaned on GenAI to fill knowledge gaps while maintaining epistemic agency for novelty discovery. We also uncovered an expertise paradox: GenAI outputs were hardest to verify when most needed. Our empirical insights motivate GenAI designs that calibrate verification to researchers' expertise, nudge toward cross-domain synthesis, and adapt prompting and outputs to disciplinary conventions.

1 Introduction

Interdisciplinary researchers must balance depth in a primary field with acquiring knowledge in less familiar areas, often while navigating those areas independently. Although prior work studied GenAI in general or single-discipline research, its use in interdisciplinary practice remained less understood.

1 Introduction

  • Researchers must balance advancing their primary discipline with keeping pace in secondary areas, where they may have limited training.Access to domain-expert collaborators is uneven, and resource constraints can require researchers to navigate unfamiliar domains independently.Source: §1 · pp. 1–2
  • Prior GenAI research examined general or single-discipline research tasks, leaving interdisciplinary researchers’ distinct needs and practices less understood.Those needs include acquiring knowledge in less familiar domains and integrating insights across disciplines.Source: §1 · p. 2
  • The paper focuses on how interdisciplinary researchers orchestrate LLMs for knowledge sense-making as they move between areas of differing expertise.The authors position this as an extension of prior HCI work on LLM-supported research workflows [69] [127].Source: §1 · p. 2

GenAI tools to navigate challenges in their research?

The study examines how 15 interdisciplinary researchers used standard GenAI and Deep Research in their workflows. Researchers used GenAI to explore unfamiliar areas but retained responsibility for filtering and integration, while finding outputs harder to verify where their expertise was lower.

GenAI tools to navigate challenges in their research?

  • Across 15 researchers’ real-world workflows, the study triangulated 1–2 weeks of tool use with longitudinal tracking documents, interaction logs, and interviews.Participants had access to standard GenAI and Deep Research, allowing the study to examine how they selected and orchestrated the modes as needs evolved.Source: GenAI tools to navigate challenges in… · p. 2
  • Researchers mainly used GenAI for breadth-oriented exploration, retained responsibility for filtering and integration, and were less equipped to verify outputs in secondary fields.In primary fields, deeper expertise supported more critical scrutiny; the authors call this asymmetry an expertise paradox.Source: GenAI tools to navigate challenges in… · pp. 2–3
  • Researchers relied on GenAI for knowledge discovery and technical assistance in secondary areas, while novelty discovery in primary areas was harder to delegate.The study identified six challenges that motivated GenAI use and examined how those needs shaped resource allocation.Source: GenAI tools to navigate challenges in… · pp. 2–3
  • The findings motivate design support for expertise-calibrated verification, cross-disciplinary synthesis, disciplinary output conventions, and prompting in unfamiliar domains.The authors emphasize that meaningful integration still depends on researchers framing and making connections themselves.Source: GenAI tools to navigate challenges in… · p. 3

2 Related Work

Prior work describes GenAI use across research workflows and the challenges of interdisciplinary work, but offers limited understanding of how researchers navigate and integrate knowledge across disciplines. This study examines those practices and their implications for human–AI collaboration.

2.1 GenAI as Research Assistants

  • GenAI research assistants support ideation, literature review, and writing, while researchers continue to evaluate and steward outputs [119] [124].Despite these advances, the broader support of navigating and integrating knowledge across disciplines remains less understood [5] [43] [71] [132].Source: §2.1 · p. 3

2.2 Interdisciplinary Research

  • Interdisciplinary researchers balance depth in a primary field with breadth across secondary areas, yet less is known about how they navigate those commitments and use tools to do so.The paper addresses this gap by examining LLM practices for negotiating depth and breadth and shaping human–AI collaboration.Source: §2.2 · p. 4

3 Methods

The study combined a pre-interaction survey, a 1–2 week longitudinal use period, and post-interaction interviews with analysis of goals and prompts. Its methods capture tool use in participants’ ongoing research and code both goal-level choices and turn-level interaction.

3.1 Study Design

  • The three-phase study combined a pre-interaction survey, a 1–2 week interaction period, and a one-hour semi-structured interview.Participants used standard GenAI and Deep Research and documented research activities and newly acquired knowledge during the interaction phase.Source: §3.1 · pp. 4–5

3.1.1 Participant recruitment.

  • The study recruited 15 English-fluent researchers actively leading interdisciplinary projects through snowball and convenience sampling.Participants provided informed consent and received a $50 gift card.Source: §3.1.1 · p. 6

3.1.2 Phase 1: Pre-interaction Survey.

  • Participants reported stronger confidence in their primary than secondary areas: 80% rated primary confidence ≥8, compared with 26.67% for secondary confidence.The secondary-area modal confidence scores were 3 and 5.Source: §3.1.2 · p. 6

3.1.3 Phase 2: Longitudinal Interaction Period.

  • The longitudinal phase collected 411 prompt–response pairs, 27% involving Deep Research, and participants rated each goal’s perceived success.Participants averaged 27.4 turns and 6.27 goals each; success ratings reflected perceived advancement, not independently measured output accuracy.Source: §3.1.3 · p. 7

3.1.4 Phase 3: Post-Interaction Interview.

  • Post-period interviews elicited participants’ reasoning about goals, disciplinary areas, and the relative strengths and shortcomings of the two modes.Interview accounts were triangulated with interaction logs to examine whether reported strategies appeared in tool use.Source: §3.1.4 · p. 7

3.2 Data Analysis

  • The analysis examined transcripts, tracking documents, and interaction logs at both goal and prompt levels.Goal-level coding captured usage type and disciplinary area; prompt-level coding characterized how participants operationalized research goals.Source: §3.2 · p. 7

3.2.1 Goal-Level Analysis.

  • The goal-level analysis classified 69 goals by disciplinary relevance and task type, with technical information seeking (n=26) and paper discovery (n=13) most common.Programming and brainstorming each accounted for n=9, and writing for n=7; the remaining four categories were comparatively rare.Source: §3.2.1 · p. 8

3.2.2 Prompt-Level Analysis.

  • The prompt-level scheme organized 411 tracked turns into five themes, allowing prompts to receive multiple codes and disciplinary relevance labels.The themes were Grounding, Information-seeking, Iterative refinement, Knowledge Synthesis, and Miscellaneous; the scheme yielded 678 code instances. Two authors independently coded four participants’ conversations with 95% percentage agreement.Source: §3.2.2 · p. 8

4 Findings

The findings map six needs behind researchers’ use of GenAI, from filling secondary-domain knowledge and method gaps to integrating perspectives and pursuing novelty. Researchers matched tools to goals and actively steered and checked outputs, with the interaction patterns varying by disciplinary context.

4.1.1 Challenge 1: Knowledge Discovery.

  • Knowledge discovery was common (N=11), especially in secondary areas where researchers needed to build background quickly and struggled to verify unfamiliar information.Challenges included finding cross-field search vocabulary and locating suitable infrastructure, not simply accessing information.Source: §4.1.1 · p. 9

4.1.2 Challenge 2: Integration.

  • Integration required connecting concepts or methods across domains, and researchers—not Deep Research alone—had to frame prompts to elicit those connections.Participants described forced or unintuitive links when the tool tried to bridge domains without explicit framing.Source: §4.1.2 · p. 10, §4.2.1 · p. 13

4.1.3 Challenge 3: Method.

  • Five participants faced secondary-domain method or skill gaps, using GenAI both to carry out tasks and to learn unfamiliar field practices.Framing also required interpreting concepts through another field’s perspective and using terminology that made research legible at the disciplinary intersection.Source: §4.1.3 · p. 10, §4.1.4 · pp. 10–11

4.1.5 Challenge 5: Collaboration.

  • Researchers had difficulty accessing collaborators with needed expertise, while GenAI could offer guidance when relevant human expertise was unavailable or hard to reach.Novelty challenges instead arose in participants’ primary areas, where little precedent existed for GenAI to draw on and complex new work was difficult to delegate.Source: §4.1.5 · p. 11, §4.1.6 · p. 11

4.2.1 GenAI Allocation for Interdisciplinary Research Challenges.

  • For challenges tied to knowledge discovery and methodology, participants often used Deep Research; all three novelty-related goals were rated Not Successful regardless of mode.Deep Research could surface papers, keywords, and methodological justifications, but its broader search offered no advantage when no prior work existed to retrieve.Source: §4.2.1 · pp. 11–13

4.2.2 GenAI Allocation Depends on the Research Goal.

  • Deep Research was used more for technical information seeking and literature discovery, while standard GenAI was used more for programming, writing, and idea generation.These choices reflected perceived affordances: broad synthesis for information gathering versus iterative interaction for debugging, revising, and refining ideas. Outcome differences were descriptive because several goal-tool groups were small.Source: §4.2.2 · pp. 13–14

4.3.1 How does the conversation unfold turn by turn?

  • In secondary and interdisciplinary conversations, information seeking often transitioned to grounding, while those contexts also showed prominent iterative refinement.Researchers repeatedly contextualized new information and refined responses, whereas primary-area prompts showed repeated information seeking but relatively little refinement.Source: §4.3.1 · p. 14

4.3.2 Steering GenAI toward the Researcher’s intent.

  • Useful responses often required specific prompts, follow-up steering, and researchers’ expertise to filter suggestions; participants also described coarse control and confident unsupported claims.Breadth could help provide background or comparative perspectives, even when results were not a direct match for the project.Source: §4.3.2 · pp. 15–16

5 Discussion

Researchers found GenAI most useful for entering secondary fields, where it was also hardest to verify, while retaining responsibility for judgment and integration. The discussion proposes design support for verification, cross-domain synthesis, disciplinary formats, and prompting, while noting limits on the study’s scope.

5.1 The Expertise Paradox

  • 80% reported high confidence in their primary area, versus 26.67% in their secondary area, where knowledge discovery was the most common reason for seeking GenAI support.Participants described little new information in an established field but found GenAI educational in a secondary field, including when it surfaced unfamiliar concepts.Source: §5.1 · p. 16
  • The expertise paradox is that GenAI may be most valuable where researchers have less knowledge, yet they are less equipped to verify its outputs there.Participants were expert researchers entering secondary fields, so adjacent expertise may still provide some basis for evaluating unfamiliar information; researchers retain judgment over what to accept, adapt, or discard.Source: §5.1 · pp. 16–17

5.2 Implications for Design

  • For secondary-domain verification, systems should express uncertainty and disclose when relevant materials cannot be found rather than return irrelevant or exaggerated responses.The discussion also proposes translating unfamiliar findings’ methods and evidentiary structures into terms researchers know, while flagging when those methods are standard in the secondary field.Source: §5.2 · pp. 17–18
  • Interdisciplinary tools should support synthesis across domains with separate modes for broad information seeking and deeper integration of researcher-curated information.The authors report that synthesis rarely sustained itself across turns, unlike information seeking, iterative refinement, and grounding.Source: §5.2 · p. 18
  • Output formats could adapt to disciplinary conventions, such as diagrams, slides, demo videos, code, maps, or archival resources.Participants in mixed reality and American history and memory studies described different format needs.Source: §5.2 · p. 18
  • Prompting support could recommend domain-relevant vocabulary and framing, or translate researchers’ familiar questioning patterns into well-formed questions for secondary fields.Participants also suggested exchanging prompting strategies tailored to their domains through a community of interdisciplinary researchers.Source: §5.2 · pp. 18–19

5.3 Limitations

  • The study’s sample of 15 limits generalizability, its two-week duration captures only initial interactions, and the free Deep Research version was restricted to five uses at the time.Success ratings measured perceived helpfulness, not verified correctness, by design.Source: §5.3 · p. 19

6 Conclusion

The conclusion describes interdisciplinary researchers using GenAI mainly for information foraging while shifting between expert, learner, and integrator roles. The participant interest form collects information about research areas and project involvement.

6 Conclusion

  • Researchers treated GenAI mainly as an information-foraging assistant, shifting between expert, learner, and integrator roles as disciplinary context changed.They sometimes tolerated irrelevant or imprecise outputs when these offered useful directions, inspiration, broader perspectives, or cross-field connections.Source: §6 · p. 19
  • Researchers still evaluate and integrate unfamiliar knowledge into their research context, so future assistants should preserve epistemic agency rather than replace human expertise.The conclusion calls for multiple modes of engagement and flexible granularity of control.Source: §6 · p. 19

A Participant Interest Form

  • The participant interest form asks about applicants’ majors, research areas, interdisciplinary project leadership, and second research area.It also asks whether applicants are fluent in English and how many team members are in their project.Source: A Participant Interest Form · pp. 24–25
  • The form distinguishes current or planned interdisciplinary projects from non-interdisciplinary projects and asks about willingness to share research progress.The listed project-status options also include not leading a research project.Source: A Participant Interest Form · p. 25

B Pre-interaction Survey

The pre-interaction survey asks about researchers’ fields, confidence, projects, research stage, and familiarity with GenAI tools. A separate subsection asks respondents about their attitudes toward AI’s effects.

B Pre-interaction Survey

  • The survey asks respondents to identify their first and second research areas and rate confidence in their knowledge for research needs in each.It also asks for a project title and description and the respondent’s current research stage.Source: B Pre-interaction Survey · p. 26
  • Respondents report their familiarity with GenAI, technical LLM research, and deep research tools, including which deep research tools they have used.Source: B Pre-interaction Survey · p. 26

B.1 Attitude toward Artificial Intelligence

  • The AI-attitudes subsection asks respondents to rate whether AI will destroy or benefit humankind and whether it will cause many job losses.Source: §B.1 · p. 26

B.2 Other questions

The appendix asks respondents for demographic and institutional information and includes a figure labeled as responses to a question about first research area.

B.2 Other questions

  • The questionnaire asks respondents for their gender, age range, institution, and academic level.Listed academic levels include junior and senior PhD students and junior and senior industry researchers.Source: §B.2 · pp. 26–27
  • Figure 8 is labeled as responses to the question, “What is your first research area?”Source: Figure 8 · p. 27

C Participant statistics

The section points to pre-interaction survey responses and tables describing prompt labels, their descriptions, and examples.

C Participant statistics

D Prompt

  • Table 6 gives prompt-label descriptions and examples for Grounding, Information Seeking, and Iterative Refinement.Source: D Prompt · p. 27
  • Table 7 gives prompt-label descriptions and examples for Knowledge Synthesis and Miscellaneous.Source: D Prompt · p. 27

E Goal labeling

This section defines goal-type labels and points to tables reporting prompt-label frequencies and goal- and prompt-level statistics by disciplinary context.

E Goal labeling

  • Discover papers denotes literature searching, writing covers writing assistance, and programming concerns coding or programming problem-solving.Source: E Goal labeling · pp. 28, 32, 33
  • Brainstorming covers idea generation, while summarization condenses information or extracts key points from longer texts or papers.Source: E Goal labeling · p. 34
  • Suggestion is advice-seeking, and proof is verification or validation of information.Source: E Goal labeling · p. 34
  • Data analysis covers requests to analyze data, while technical information seeking concerns non-programming technical solutions.Source: E Goal labeling · pp. 27, 34

F Prompt and goal label statistics

G Normalization for Participant Imbalance

Because turn and transition counts varied sharply across participants, the authors normalized transition frequencies within each participant before aggregating them equally.

G Normalization for Participant Imbalance

  • Two participants contributed 43.3% of 411 turns and 53.2% of 1,182 transitions, so pooled counts could disproportionately reflect their conversational habits.Source: G Normalization for Participant… · p. 34
  • The method computes each participant’s proportion for each transition, then averages those proportions with equal participant weight; display rescaling changes edge widths only.Source: G Normalization for Participant… · pp. 34, 36
  • Normalization shifts weight away from transitions driven mainly by one or two high-volume participants and toward transitions recurring across participants.Source: G Normalization for Participant… · p. 36
  • Self-loop transitions fell from 104 raw transitions to 62.8 after normalization, indicating refinement-on-refinement was disproportionately characteristic of two high-volume participants.Source: G Normalization for Participant… · p. 37
Loading 2609.30588v1…