Nonparametric domain adaptation for multimodal connoisseurship of Buddhist statuary across dynasties (ChronoStyleNet2.0)

Technical Summary: ChronoStyleNet 2.0 (CSN 2.0)

Problem Statement
The documentation, dating, and interpretation of Chinese Buddhist statuary face significant constraints due to a shortage of trained specialists, the vast scale of undocumented collections, and urgent conservation needs. Traditional reliance on specialist visual examination and textual research is difficult to scale. While digital methods (2D deep learning, 3D point-cloud analysis) and general-purpose Multimodal Large Language Models (MLLMs) offer new avenues, they often lack specific archaeological terminology, context, or the ability to provide interpretable pathways from visual evidence to stylistic and chronological attribution. Previous attempts at domain adaptation (e.g., ChronoStyleNet 1.0) faced limitations regarding chronological coverage, the risks of parameter fine-tuning, and "premature attribution" where injected knowledge overrode ambiguous visual evidence.

Methodology: ChronoStyleNet 2.0 (CSN 2.0)
CSN 2.0 (Chinese name: Zhijian Zaoxiang) is a nonparametric domain adaptation system designed to broaden the chronological and geographic scope of Buddhist statuary analysis without further parameter fine-tuning of the underlying model.

• Core Architecture: The system utilizes Gemini 2.5 Flash as the multimodal reasoning backbone. It operates via a four-stage inference workflow: query parsing, context assembly, model inference, and response parsing.
• Nonparametric Adaptation Strategy: Instead of fine-tuning, CSN 2.0 injects domain knowledge at inference time through three integrated components:
  1. Structured Prompting: A ~4,000-word domain prompt containing a nine-step analysis protocol, dynastic feature summaries, and an evidence-priority scheme. Crucially, it enforces a "visual-evidence-first" rule, requiring the model to describe observable features before making iconographic or chronological attributions.
  2. Tri-Strategy Fusion Retrieval (TSFR): A retrieval-augmented generation (RAG) module that uses relational, text-based retrieval rather than vector databases. TSFR combines:
     • Direct full-text matching (SQL LIKE) for standardized terms.
     • Tokenized keyword retrieval for fragmented queries.
     • Curated domain mapping to expand searches based on trigger terms.
  3. CIDOC CRM-Informed Domain Ontology: A knowledge base organized into 11 layers (e.g., Temporal, Spatial, Iconographic, Material) mapped to selected CIDOC CRM classes and properties. This structure links visible features to reference materials and proposed attributions, though the implementation is described as "informed" rather than strictly compliant due to simplifications (e.g., omitting event-centered modeling).
• Datasets:
  • Knowledge Base: 3,642 curated entries spanning the Sixteen Kingdoms to the Qing Dynasty, including iconographic references, local gazetteers, and archaeological literature.
  • Structured Statue Dataset: 145 annotated statues across nine iconographic categories (e.g., Śākyamuni, Avalokiteśvara) from various dynasties.
  • Evaluation Set: A held-out set of 18 statues (two per iconographic category) covering nine dynasties, used to test the system against four general-purpose models.

Key Contributions

1. Dataset Creation: Development of an expert-annotated, multi-dynasty dataset of Buddhist statuary and a separate 18-item evaluation set.
2. Nonparametric Workflow: Implementation of a domain adaptation workflow combining structured prompting, TSFR, and a CIDOC CRM-informed ontology, avoiding the risks of parameter fine-tuning.
3. Prompt-Level Mitigation: Introduction of a "visual-evidence-first" rule designed to reduce premature attribution by forcing the model to link conclusions to observable features before finalizing dates or identities.

Results
The system was evaluated on 18 held-out sculptures across eight archaeological dimensions (e.g., stylistic feature extraction, dating accuracy, cultural interpretation, reasoning soundness) and compared against GPT-5.1, Claude 4.5, Gemini 2.5 Pro, and Grok 3.

• Performance Strengths: CSN 2.0 achieved relatively high scores in stylistic feature extraction and linguistic consistency compared to other non-GPT models. It showed mixed performance in dating accuracy (mean score 2.944), outperforming some general models but scoring significantly lower than GPT-5.1 (4.944). Performance in statue-type recognition was heterogeneous: the system scored 4 or 5 on Avalokiteśvara, Samantabhadra, and Mañjuśrī items, but received scores of 1 or 2 for Maitreya and Mahāsthāmaprāpta items, indicating substantial variability across categories.
• Performance Weaknesses: The system scored lower in cultural interpretation and evidence-to-conclusion reasoning compared to GPT-5.1 and, in some dimensions, Gemini 2.5 Pro.
• Error Analysis: While CSN 2.0 provided detailed anatomical descriptions, it occasionally failed to identify the specific diagnostic feature required for correct attribution (e.g., misinterpreting a treasure vase as a Buddha image). High standard deviations in scores reflected heterogeneity across the test items rather than run-to-run instability.
• Comparative Context: GPT-5.1 received the highest mean scores across all dimensions. CSN 2.0 generally outperformed Claude 4.5 and Grok 3 but showed mixed results against Gemini 2.5 Pro depending on the specific dimension.

Significance and Claims
The paper positions CSN 2.0 not as a replacement for expert judgment, but as an exploratory, traceable aid for specialist analysis.

• Feasibility of Nonparametric Adaptation: The study demonstrates the practical feasibility of supplying a general-purpose multimodal backbone with curated domain information at inference time. This approach allows the knowledge source to be updated independently of the model and makes retrieved evidence inspectable.
• Traceability: By using structured prompts and a domain ontology, the system aims to make the reasoning path from visual evidence to attribution more transparent, addressing the "black box" nature of general MLLMs.
• Limitations and Scope: The authors explicitly state that the evaluation is exploratory. Due to the small test set (18 items), single-run outputs, and the lack of component ablation studies, the results cannot be attributed to specific components (e.g., the ontology vs. the prompt) nor generalized broadly. The system is intended for preliminary documentation, teaching, and hypothesis generation, with outputs requiring specialist verification. The authors caution against using the system for conservation decisions without further validation.