Source-linked AI summary
Sanyu Studio: A Multi-Agent System for Art-Historical Narrative Construction
Zhaoxi Wei, Hongye Yang, Shuyuan Tian
TL;DR
Art-historical narrative formation with LLMs remains underexplored amid concerns that statistical generation may flatten artistic meaning. Sanyu Studio models paintings as evidentiary agents in a roundtable dialogue, and the study finds that this process broadens participation and produces divergent yet coherent narratives.
Problem
How LLM systems can support plural art-historical narrative formation remains underexplored amid concerns that statistical generation may homogenize artistic meaning.
Method
Sanyu Studio proceduralizes art-historical inquiry through painting agents, user questions, organized evidence, and roundtable interpretive negotiation.
Results
With DSBM enabled, high-information works’ speaking share fell from 41.8% to 20.4%, while distinct works discussed increased from 151 to 279, broadening narrative participation.
Takeaways & Limitations
Sanyu Studio offers nonexpert audiences an accessible interactive entry into art-historical interpretation through comparison and negotiation of evidence and interpretive paths.
Takeaways & Limitations
The system may over-infer when addressing personal experience, private relationships, or sensitive identity, creating ethical and privacy risks.
Abstract
from arXiv · showhide
Amid concerns that generative AI may standardize art interpretation, this paper examines whether LLM-based interaction can support plural art-historical narrative construction. We present Sanyu Studio, a multi-agent dialogue system that models 321 Sanyu oil paintings as agents with fact, interpretation, organization, and memory-filtering mechanisms. Based on a seven-day workshop with eight art-university participants, the study shows that user prompts, evidence organization, and cognitive tendencies shaped divergent yet coherent versions of digital Sanyu. The findings suggest that, under conditions of limited historical evidence, AI can amplify human agency and offer public audiences an interactive entry point into art-historical interpretation.
1 INTRODUCTION
The paper argues that a designed multi-agent LLM system can simulate art-historical narrative formation rather than homogenize artistic meaning. It presents Sanyu Studio as an interactive system that makes narrative formation visible and supports more plural artistic meanings through evidence, interpretation, and user questioning.
- LLM-mediated art interpretation raises concern that statistical generation may homogenize and flatten artistic meaning.
- A designed multi-agent LLM system can simulate the formation of art-historical narrative.
- Under constrained system design, AI can serve nonexpert researchers as a narrative mechanism enabling more plural forms of artistic meaning.
- Existing LLM-based art-interpretation approaches emphasize static image analysis or knowledge retrieval, while historical-figure chatbots use fixed personas and overlook user-shaped interpretive paths.
- Sanyu Studio maps research questions, evidence retrieval, material negotiation, and narrative synthesis onto an interactive multi-agent mechanism.
- The system makes art-historical narrative formation visible through questioning, selection, interpretation, and synthesis, using Sanyu’s open scholarly reinterpretation as its case.
2 METHODOLOGY
Sanyu Studio turns user questions into research questions that retrieve painting-based evidentiary standpoints, coordinate their negotiation, and filter exchanges into a continuously revisable image of Sanyu. Its 321 painting agents combine constrained factual and interpretive knowledge with document-scarcity balancing and memory mechanisms to broaden participation and preserve evolving narrative paths.
- System architecture: User questions function as research questions, determining which painting agents are retrieved and initiating a roundtable discussion that organizes art-historical materials around each inquiry.The system’s Sanyu agent records, filters, and synthesizes these exchanges into a revisable narrative.
- Painting agents: 321 Sanyu oil paintings are modeled as separate agents whose responses are constrained by metadata, visual features, related literature, and distinct factual and interpretive knowledge layers.Factual entries cover catalogued and verifiable information, while interpretive entries derive from open-access scholarship and labeled inferences.
- Document-scarcity balancing: 41.8% to 20.4%: DSBM reduced high-information works’ token-level speaking share, while 151 to 279: distinct works discussed increased.These changes indicate broader roundtable participation and greater narrative agency for under-documented works.
- Memory and narrative synthesis: The memory system scores dialogue fragments from 1 to 10, retains fragments scoring ≥7 in long-term memory, and preserves interpretive paths while keeping factual claims tied to the fixed knowledge base.This filtering converts selection, compression, and organization into a stable yet revisable artist narrative across interactions.
3 INTERACTION AND EXPERIENCE
A seven-day workshop with eight art-university participants examined how interaction paths and cognitive tendencies shaped Sanyu agents from the same initial evidence base. The resulting narratives diverged broadly, supporting plural art-historical meaning-making, especially where evidence was limited or contested.
- Workshop Design: A seven-day offline workshop with eight senior oil-painting undergraduates tested whether shared initial conditions could produce diverse art-historical narratives.Participants were labeled A–H and began from the same system state and evidentiary base after brief training.
- Workshop Design: Each participant completed at least 50 roundtable discussions, including at least five preset-question discussions about disputed or under-documented Sanyu issues.The system recorded dialogue content, retrieval context, and memories written into each Sanyu agent.
- Interpretive Outcomes: Participants’ cognitive tendencies and question paths shaped differentiated Sanyu agents, particularly on issues with limited evidence or no clear consensus.Variation in entry points and evidence-integration paths supported plural constructions of artistic meaning.
- Narrative Divergence: Off-diagonal mean semantic similarity was low across all five questions: Q1 = 0.050, Q2 = 0.046, Q3 = 0.057, Q4 = 0.049, and Q5 = 0.114.Q5 showed slightly higher consistency, but no stable consensus emerged.
- Cognitive Tendencies: Higher-NFCS participants converged faster through fewer rounds, shorter questions, more follow-ups, and fewer topic shifts, whereas lower-NFCS participants explored more openly.Lower-NFCS participants also asked longer questions and shifted and returned between topics more frequently.
4 DISCUSSION
The discussion presents Sanyu Studio as an accessible entry into Sanyu studies and a simulation of comparative interpretive negotiation, while identifying ethical, privacy, and participant-scope limitations.
- User experience and future work: Participants agreed that Sanyu Studio offers art-history enthusiasts an accessible entry into Sanyu studies.The discussion characterized the system as useful for engaging with Sanyu-related art history.
- User experience and future work: The roundtable format effectively simulates comparison and negotiation of materials, evidence, and interpretive paths.Its value lies in structuring interaction around competing resources and interpretations.
- Limitations: The system may over-infer personal experience, private relationships, or sensitive identity, creating ethical and privacy risks despite source labeling and fact–inference separation.Future work should add stricter content boundaries and human review.
- Limitations: The workshop included only eight art-university students with basic art-historical knowledge, limiting evidence about broader usability.Future studies will recruit participants without prior training.
5 REFLECTION AND CONCLUSION
The paper presents Sanyu Studio as a multi-agent system that operationalizes art-historical narrative construction through evidence organization, interpretive negotiation, and revisable narrative generation. It argues that LLMs can support art-historical inquiry and public engagement while extending this approach to other multi-source historical domains.
- Reflection: LLMs cannot yet produce original art-historical scholarship autonomously, but they can support evidence organization, interpretive negotiation, narrative generation, and methodological mediation.This mediation can help broader audiences enter art-historical discussion and expand research’s public reach.
- System and Method: Sanyu Studio is a multi-agent roundtable system integrating painting-agent design, evidence organization, roundtable negotiation, and a human-AI workshop.The system is introduced specifically for art-historical narrative construction.
- System and Method: Using Sanyu as a case, the system organizes artwork evidence, negotiates interpretive paths, and generates revisable artist narratives.Its procedures make key research steps computational and revisable.
- Broader Implications: The approach maps question formation, evidence retrieval, multi-source negotiation, and narrative synthesis onto a computable multi-agent mechanism extendable across historical and cultural-heritage domains.The paper identifies historical figures, cultural heritage, and event history as potential application areas.