ArchGPT: Domain-Specialized GPT in Architecture & Ethics

• ArchGPT is a term for domain-specialized large-model systems that use curated data, fine-tuning, and controlled prompting to support expert workflows.
• In architectural applications, ArchGPT is applied as a multimodal model for visual reasoning that analyzes building imagery to aid education and heritage preservation.
• For ethical AI, ArchGPT serves as a developer assistant that translates legal and ethical frameworks into actionable engineering requirements.

ArchGPT is a label used in recent arXiv literature for distinct domain-specialized large-model systems rather than for a single canonical model. In one prominent usage, it denotes a multimodal architectural visual question answering model fine-tuned on a large architecture-specific image-question-answer corpus in order to analyze buildings from images with grounded, detailed visual reasoning (Wang et al., 25 Sep 2025). In another, it denotes a proposed custom developer GPT intended to help AI engineers translate legal requirements and ethical concerns into engineering decisions, especially under frameworks such as the EU AI Act and GDPR (Olson, 2024). Across these usages, the shared premise is that general-purpose models are often too broad, too weakly grounded, or too difficult to control for expert workflows, and that domain specialization must therefore be introduced through curated data, domain pretraining, retrieval, or controlled prompting.

1. Scope of the term and nearby usages

The term is best understood as polysemous. In the architecture-and-vision literature, ArchGPT is a model for architecture-specific multimodal understanding from images. In AI ethics and software engineering, ArchGPT is a developer-facing conversational artefact for ethical AI solutions. A closely related naming variant, ArcGPT, refers to a general-purpose LLM tailored to the archival field, where the central tasks are retention period prediction, open-access identification, confidentiality prediction, and post-OCR processing rather than architectural interpretation or ethical design support (Zhang et al., 2023).

This naming overlap is not accidental. Each system is positioned against a similar failure mode of generic LLMs. Architectural multimodal models are described as vulnerable to language-prior bias and visual hallucination in architectural settings. Developer-facing ethics assistants are motivated by the difficulty of turning general ethical and legal principles into concrete software features. Archival systems are motivated by historical language, OCR noise, archival terminology, and workflow-specific decisions that are poorly captured by general text models. This suggests a broader research pattern in which “GPT”-style systems are adapted not by expanding generality, but by constraining the model to a narrower operational domain.

2. ArchGPT as a multimodal model for architectural understanding

In the architectural interpretation setting, ArchGPT is a domain-adapted multimodal model trained to answer architecture-specific questions about images and to support architecture education, heritage preservation, creative workflows, and immersive exploration. The model is produced by supervised fine-tuning of ShareGPT4V-7B in a LLaVA-style setup on Arch-300K, an architecture-focused corpus of 315,247 image-question-answer triplets covering 8,643 unique architectural scenes, with up to 8 images per scene and an average of 72.9 questions per scene (Wang et al., 25 Sep 2025).

The model stack is explicitly modular.

Component	Specification	Role
Dataset	Arch-300K, 315,247 triplets	Architecture-specific supervision
Vision encoder	CLIP ViT-L/14, $336 \times 336$, 576 visual tokens	Visual feature extraction
Projection layer	Two-layer MLP	Maps visual tokens to LLM space
LLM	Vicuna-v1.5 7B, based on LLaMA-2	Text generation and reasoning
Fine-tuning method	LoRA, $r=128$, $\alpha=256$	Parameter-efficient adaptation

The supervised fine-tuning configuration is also specified: two epochs, batch size 16, learning rate $2 \times 10^{-5}$, maximum context length 2048, AdamW, cosine schedule, 3% warmup, weight decay 0.0, bfloat16 and TF32, gradient checkpointing, and DeepSpeed ZeRO-3 on a 40GB single GPU. The paper describes a multi-annotation fusion strategy in which all annotations for one image are concatenated into a single training sequence with attention masks, thereby avoiding repeated image encoding and reducing training time by more than half. The training objective is standard next-token prediction on image-conditioned instruction-response sequences.

The model is not presented as a new multimodal backbone. Its contribution lies in domain adaptation: using a curated architecture-specific dataset, filtering images for architectural clarity, refining textual metadata, and fine-tuning an open multimodal backbone so that long-form answers become more faithful to architectural evidence. The paper’s intended application modes are an interactive architectural conversational assistant, AR-enabled architectural interpretation, and VR virtual walkthroughs with live question answering.

3. Arch-300K and the curation pipeline

The data-construction pipeline is central to the architectural version of ArchGPT. Arch-300K is built through five stages: scene-level architectural scene selection, coarse-to-fine image filtering, LLM-guided text verification and knowledge distillation, formal analysis annotation synthesis, and manual screening and refinement (Wang et al., 25 Sep 2025).

Architectural scenes are drawn primarily from Wikimedia Commons. Scene selection uses architectural and non-architectural keyword dictionaries over scene names and taxonomy, with Gemini 2.5 Pro used when a scene is ambiguous. Roughly 40% of candidate scenes are classified as architectural at this stage. The image curation stage then addresses the fact that Wikimedia scenes contain unconstrained tourist photo collections with varying viewpoints, weather, cameras, and occlusions. The paper uses a coarse-to-fine strategy that combines 3D reconstruction confidence and semantic segmentation. VGGT is used as a feed-forward 3D reconstruction model to estimate structural consistency, with a global threshold based on the $\alpha$-quantile of pixel confidence values and $\alpha=0.8$. SAM then refines the resulting architectural masks, and the top $K$ images are retained with $K=8$. Approximately 55% of architectural scenes are discarded entirely because no qualifying images remain after filtering.

Text supervision is refined through two additional mechanisms. First, an “LLM-known checker” asks whether the model knows the architecture; if the answer is NO, the scene is excluded from text annotation. Second, a Wiki-data extractor combines the model’s prior knowledge with raw Wikimedia metadata to extract a refined architectural description, the geographical location and construction year, and the canonical formal architecture name. Extraction is constrained to the provided metadata, and missing information is emitted as “Unknown.” The paper explicitly describes this as a retrieval-augmented generation-like process in which raw wiki metadata acts as retrieved context.

The final annotations comprise two complementary supervision types. Detailed descriptions are paragraph-level narratives that summarize the visual character of a building while avoiding direct naming. Aspect-guided conversations provide targeted question-answer pairs over categories such as architectural style, architectural elements, architectural details, architectural context, architectural innovation, architectural symbolism, and architectural materials. About 23% of the questions are detailed-description prompts and 77% are aspect-guided conversations. Answers are long-form, averaging 81.7 words. Ablation results reported in the paper indicate that the LLM-known checker, Wiki-data extractor, and coarse-to-fine image filter each contribute to better performance, with the image filter yielding the best results among the tested curation components.

4. ArchGPT as a developer assistant for ethical AI solutions

A separate paper uses the name ArchGPT for a custom developer GPT intended to help AI engineers discuss and solve ethical issues through AI engineering. Its stated functions are legal and compliance guidance, especially for the EU AI Act and GDPR, and ethical perspective expansion through concerns drawn from minoritized communities (Olson, 2024).

This system is framed not as a finished deployed model, but as a conceptual and implementation-oriented proposal. The paper characterizes ArchGPT as both an “interactive, data-driven user story” and a “values translator,” with the latter defined as a tool that communicates values concepts in terminology understandable for development teams. The intended users are AI engineers and software developers engaged in epics and user stories, feature prioritization, sprint planning, design, testing, and general AI engineering decisions.

The proposed implementation has three steps. First, it builds a dataset of ethical concerns; the paper states that earlier work already collected over 2000 Reddit posts from seven minoritized communities and manually annotated them for software-related ethical concerns. Second, it explores GPT bias and safeguard behavior through custom prompts in order to study refusal patterns, latent bias, and distortions in responses about ethical harms. Third, it creates the custom GPT by combining the collected dataset, recent legal and ethical frameworks, deterministic responses for top concerns, and prompt-based interaction. The paper explicitly argues that deterministic responses are needed for critical questions because GPTs are non-deterministic and may have unknown biases and safeguards.

The worked example is non-consensual pornography. The paper treats it as a cyberbullying and safety problem, notes that in several countries it is a crime, and states that platforms may be legally accountable if they do not take reasonable measures to remove it. The cited workflow is that a developer asks for women’s ethical concerns about social platforms, narrows the issue to non-consensual pornography, and then asks what legal requirements and feature directions apply. The envisioned response is a combination of real women’s ethical complaints, relevant legal requirements, and candidate AI-centered mitigation features such as automatically identifying and removing non-consensual pornography. The paper also reports a background statistic from cited prior work: nearly 40% of platforms lacked reporting interfaces, and only 16% allowed users to report using appropriate legal vocabulary. No completed quantitative or qualitative evaluation of ArchGPT itself is reported; user studies with AI engineers are left as future work.

5. Relation to adjacent domain-specialized GPT systems

The broader domain-specialization pattern surrounding ArchGPT is visible in closely related systems such as ArcGPT for archives and AMGPT for metal additive manufacturing.

System	Domain	Core adaptation
ArcGPT	Archival applications	7B BatGPT-based model further pretrained on archival-domain data
AMGPT	Metal additive manufacturing	Llama2-7B in a RAG setup over about 50 AM papers and textbooks
ArchGPT	Architectural VQA	ShareGPT4V-7B fine-tuned on Arch-300K
ArchGPT	Ethical AI engineering	Custom GPT proposal grounded in ethical concerns and legal frameworks

ArcGPT is presented as the first general-purpose LLM tailored to the archival field. It is a 7B model built on the BatGPT architecture and further pretrained on broad archival-domain data spanning archival journals, archival records, and documents from multiple historical periods. Its benchmark, AMBLE, comprises four real-world tasks: retention period prediction, open-access identification, confidentiality prediction, and post-OCR processing. On AMBLE, ArcGPT is the best among the reported generative LLM baselines on the classification tasks, reaching 84.40 F1 on retention period prediction, 84.00 F1 on open-access identification, and 94.40 F1 on confidentiality prediction, but it does not exceed the strongest discriminative baselines on every task and performs substantially worse than dedicated spelling-correction models on post-OCR processing, where Mengzi-T5-Base-csc reaches 10.90 Levenshtein Distance against ArcGPT’s 38.86 (Zhang et al., 2023).

AMGPT is a domain-specific assistant for metal additive manufacturing implemented as Retrieval-Augmented Generation rather than from-scratch pretraining. Its workflow is explicit: the user asks a question, the query is embedded, cosine similarity is used to retrieve semantically relevant chunks from a vector database, the retrieved text is concatenated with the prompt, the LLM generates an answer conditioned on both query and context, and a chat-memory layer maintains coherence across turns. The system uses a pre-trained Llama2-7B model, Mathpix PDF-to-TeX conversion, and LlamaIndex as the RAG infrastructure layer over about 50 AM papers and textbooks. In the reported comparison set, the RAG system is factual on 80% of prompts, whereas GPT-4 is factual on 86.7% of prompts, but the paper also states that the RAG system can remain more specific than GPT-4 on certain AM questions and that expert evaluations found specific embeddings from the RAG setup accelerated response times and maintained coherence (Chandrasekhar et al., 2024).

These neighboring systems show that ArchGPT is part of a wider methodological shift toward expert models whose value is not reducible to raw open-domain accuracy. Domain pretraining, retrieval over curated corpora, structured metadata extraction, and benchmark construction recur across archives, manufacturing, architecture, and ethics-oriented tooling.

6. Methodological themes, limitations, and open questions

Taken together, the ArchGPT-related literature suggests a recurring transition from unconstrained chat toward structured expert systems. The architectural ArchGPT relies on high-quality image filtering, metadata verification, and supervised fine-tuning on a domain-specific corpus. The ethical ArchGPT proposes deterministic responses, curated evidence from minoritized communities, and legal-framework awareness. ArcGPT uses domain pretraining over archival text, while AMGPT uses retrieval over a fixed technical corpus. This suggests that “specialization” in these works is achieved through different control points—data curation, retrieval, fine-tuning, or prompt design—rather than through a single recipe (Wang et al., 25 Sep 2025).

A common misconception is that a domain-specialized GPT necessarily surpasses stronger general or task-specific systems on all metrics. The reported evidence is more qualified. In archives, ArcGPT outperforms the other generative baselines on AMBLE’s classification tasks, yet still trails the strongest discriminative models on some tasks and lags far behind dedicated spelling-correction models on post-OCR cleanup. In additive manufacturing, AMGPT’s reported advantage is not raw factual accuracy over GPT-4, since the comparison cited in the paper gives GPT-4 86.7% factual prompts against AMGPT’s 80%, but rather domain-grounded specificity and controllability. In the ethical-AI setting, the ArchGPT paper does not yet provide a completed system evaluation at all, so its contribution remains a proposal rather than a demonstrated benchmark win (Zhang et al., 2023).

The limitations are correspondingly domain-specific. The architectural ArchGPT excludes scenes that fail the LLM-known check and discards a large fraction of candidate imagery during filtering, which leaves open the question of robustness on obscure buildings or scenes with weak metadata. The ethical ArchGPT depends on a dataset derived from Reddit posts from seven minoritized communities, which is valuable but not exhaustive, and its usability with real development teams remains untested. ArcGPT’s benchmark is entirely Chinese and based on collaboration with one administrative archive unit. AMGPT is limited by a corpus of about 50 documents, by retrieval quality, by context length, and by inference parameters such as temperature, max token length, and top-$k$ retrieval, with the paper indicating that lower temperature and smaller top-$k$ worked best for factual AM question answering (Chandrasekhar et al., 2024).

The resulting research agenda is therefore not simply to scale model size. The papers point instead toward benchmarked domain adaptation, stronger curation pipelines, more explicit grounding, and closer integration with operational workflows. In that sense, ArchGPT names a class of specialist systems whose defining problem is not whether a LLM can speak fluently, but whether it can be made faithful, controllable, and procedurally useful inside a domain that already has its own evidentiary standards, task structure, and professional judgment.