Kaicong Wu

Kaicong Wu

orcid

Research Focus & Metrics

Academic Career Length

11Years (2016-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications

1 January 2018
Verified
Conference Paper
Closed access
PDF file
KWKaicong Wu
Kaicong Wu
AKAxel Kilian
Axel Kilian

29 August 2019
Verified
Book Chapter
Closed access
PDF file
KWKaicong Wu
Kaicong Wu
AKAxel Kilian
Axel Kilian


15 September 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
ZLZihang Liu
Zihang Liu
HYHua Yu
Hua Yu
HZHaotian Zhang
Haotian Zhang
KWKaicong Wu
Kaicong Wu
The built environment fundamentally shapes everyday experience, well-being, and social interaction, yet the capacity to critically evaluate design solutions remains difficult to transfer and access publicly because it requires the integration of visual evidence, tacit knowledge, domain expertise, and professional judgment, capacities that remain inadequately supported by current AI systems. To address this gap, we introduce a domain-specific pipeline that repositions large language models (LLMs), particularly vision-language models (VLMs), as design critics by combining domain fine-tuning, structured critique supervision, and tacit knowledge modeling for the evaluation of built-environment design solutions. We constructed a curated architectural reference corpus, integrating visual exemplars with expert commentary to preserve semantic coherence. Representative cases were constructed as image-text pairs and analyzed using pretrained LLMs to generate multi-perspective question-answer (Q&A) training pairs reflecting expert-style critique. We then fine-tuned an open-source VLM, Qwen3-VL, using low-rank adaptation (LoRA), a parameter-efficient strategy applied to multimodal embeddings. Meanwhile, we developed an attention-based mechanism, combining multilayer perceptrons (MLPs), that learns task-specific weightings over tacit knowledge dimensions spanning psychological, spatial, aesthetic, structural, and tectonic criteria, enabling the model to associate design variations with corresponding evaluative emphasis. Experiments show that our pipeline enables the VLM to approach the performance of closed-source LLMs, such as GPT-5-mini, on built-environment-specific evaluation tasks, yielding more consistent, context-sensitive, and interpretable critiques. Beyond supporting education and professional practice, our work demonstrates an effective approach to deploying LLMs as evaluative agents in domains where expert judgment relies on tacit, multidimensional knowledge and where public understanding is critical to informed decision-making.

8 September 2026
Verified
Book Chapter
Closed access
PDF file
YJYilun Jiang
Yilun Jiang
DMDaniel Chad McKee
Daniel Chad McKee
KWKaicong Wu
Kaicong Wu