CiQi-Agent: Aligning Vision, Tools and Aesthetics in Multimodal Agent for Cultural Reasoning on Chinese Porcelains

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• CiQi-Agent (7B) outperforms GPT-5 by an average of 12.2% in accuracy across six porcelain connoisseurship attributes on the CiQi-Bench benchmark.result 95%
• The CiQi-VQA dataset contains 29,596 porcelain specimens, 51,553 images, and 557,940 visual question-answering pairs.technical 95%
• CiQi-Agent is trained using a two-phase iterative framework combining supervised fine-tuning (SFT) and reinforcement learning (RL) with a tool-augmented reasoning framework.method 90%
• The model's reasoning framework integrates a vision tool for image zoom-in and two multimodal retrieval tools (image and text retrieval).technical 90%
• The effectiveness of CiQi-Agent depends on the quality and comprehensiveness of its training dataset.limitation 85%
• The CiQi-Bench evaluation benchmark comprises 775 high-quality porcelain specimens assessed via multiple-choice and free-form generation protocols.technical 90%
• The tool could replace the need for in-depth expert analysis, speeding up authentication and reducing costs for institutions managing large porcelain collections.market 80%
• The model uses a Group Relative Policy Optimization (GRPO) based RL stage with a strong tool-calling reward to learn strategic tool invocation.method 85%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card