Datasets:

MarkrAI

/

AutoRAG-evaluation-2024-LLM-paper-v1

like 8

bee1d367-b88b-43c0-a994-a09f6282794d

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## E Details Of Training And Evaluation Datasets .0 30.0 46.0 46.0 46.0 46.0 42.0 40.0 36.0 38.0 14.0 10.0 COPA 76.0 74.0 68.0 74.0 74.0 78.0 76.0 72.0 80.0 80.0 76.0 22.0 60.0 HellaSwag 58.0 30.0 36.0 34.0 28.0 52.0 44.0 50.0 46.0 46.0 38.0 16.0 2.0 sentiment SST-2 98.0 98.0 98.0 86.0 86.0 98.0 100.0 98.0 96.0 98.0 96.0 82.0 26.0 Yelp 98.0 98.0 100.0 94.0 92.0 98.0 98.0 98.0 98.0 98.0 98.0 82.0 0.0 IMDB 96.0 98.0 98.0 88.0 82.0 98.0 98.0 100.0 100.0 100.0 98.0 90.0 4.0 sentiment140 68.0 68.0 68.0 80.0 74.0 72.0 70.0 64.0 64.0 64.0 64.0 64.0 14.0 READING Comp. MultiRC 88.0 72.0 36.0 64.0 42.0 66.0 44.0 44.0 48.0 40.0 40.0 72.0 2.0 SQuADv1 74.0 62.0 62.0 58.0 58.0 60.0 60.0 64.0 66.0 64.0 60.0 42.0 8.0 SQuADv2 66.

f10b7325-7426-43c2-b9cb-8b4117bd0921

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## E Details Of Training And Evaluation Datasets 0 READING Comp. MultiRC 88.0 72.0 36.0 64.0 42.0 66.0 44.0 44.0 48.0 40.0 40.0 72.0 2.0 SQuADv1 74.0 62.0 62.0 58.0 58.0 60.0 60.0 64.0 66.0 64.0 60.0 42.0 8.0 SQuADv2 66.0 58.0 34.0 38.0 24.0 34.0 24.0 26.0 22.0 24.0 24.0 36.0 0.0 OBQA 86.0 80.0 72.0 76.0 76.0 88.0 82.0 80.0 76.0 78.0 76.0 48.0 0.0 BoolQ 84.0 68.0 66.0 84.0 82.0 78.0 78.0 86.0 84.0 88.0 86.0 74.0 10.0 Drpo 58.0 22.0 18.0 20.0 14.0 36.0 20.0 22.0 24.0 20.0 22.0 20.0 0.0 CLOSE-BOOK QA NQ 30.0 28.0 12.0 22.0 14.0 24.0 16.0 12.0 10.0 12.0 10.0 6.0 2.0 ARC-e 90.0 90.0 88.0 92.0 92.0 94.0 94.0 90.0 90.0 86.0 90.0 58.0 8.0 ARC-c 68.0 66.0 58.0 64.0 60.0 68.0 70.0 68.0 60.0 68.0 62.0 38.

5620249f-4ad0-4f23-ab97-71072b455762

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## E Details Of Training And Evaluation Datasets 0 10.0 12.0 10.0 6.0 2.0 ARC-e 90.0 90.0 88.0 92.0 92.0 94.0 94.0 90.0 90.0 86.0 90.0 58.0 8.0 ARC-c 68.0 66.0 58.0 64.0 60.0 68.0 70.0 68.0 60.0 68.0 62.0 38.0 4.0 TriviaQa 68.0 56.0 54.0 70.0 66.0 66.0 64.0 68.0 70.0 68.0 68.0 36.0 12.0 COREFERENCE DPR 90.0 90.0 72.0 68.0 66.0 88.0 72.0 76.0 76.0 68.0 72.0 52.0 20.0 WSC 58.0 60.0 58.0 42.0 52.0 64.0 56.0 46.0 48.0 44.0 42.0 58.0 0.0 READ. COOMP. W/ COMMONSENSE CosmosQa 84.0 82.0 38.0 80.0 76.0 82.0 70.0 74.0 74.0 74.0 80.0 8.0 28.0 record 80.0 78.0 28.0 34.0 22.0 74.0 46.0 28.0 22.0 24.0 18.0 18.0 0.0 PARAPHRASE Paws Wiki 92.0 70.0 52.0 74.0 50.0 88.0 74.0 54.0 66.0 54.0 62.0 90.0 8.0 QQP 90.0 88.0 76.0 66.0 56.0

d4cc5861-be27-4da4-bacd-0cc383b84614

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## E Details Of Training And Evaluation Datasets 28.0 34.0 22.0 74.0 46.0 28.0 22.0 24.0 18.0 18.0 0.0 PARAPHRASE Paws Wiki 92.0 70.0 52.0 74.0 50.0 88.0 74.0 54.0 66.0 54.0 62.0 90.0 8.0 QQP 90.0 88.0 76.0 66.0 56.0 84.0 62.0 70.0 66.0 60.0 66.0 78.0 2.0 MRPC 84.0 70.0 62.0 64.0 64.0 78.0 64.0 62.0 62.0 62.0 64.0 82.0 0.0 STSB 44.0 44.0 20.0 18.0 12.0 34.0 22.0 14.0 16.0 14.0 16.0 6.0 0.0 NLI CB 100.0 91.1 84.4 84.4 80.0 93.3 88.9 73.3 82.2 75.6 77.8 73.3 13.3 WNLI 72.0 72.0 62.0 70.0 68.0 76.0 66.0 58.0 66.0 56.0 56.0 76.0 6.0 ANLI-r1 70.0 70.0 70.0 56.0 50.0 64.0 68.0 48.0 44.0 44.0 46.0 50.0 12.0 ANLI-r2 64.0 56.0 56.0 38.0 36.0 60.0 60.0 48.0 48.0 48.0 42.0 48.0 8.0 ANLI-r3 68.0 56.0

7ea49b48-8dc3-44ba-bcb0-6204daa98cc3

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## E Details Of Training And Evaluation Datasets LI-r1 70.0 70.0 70.0 56.0 50.0 64.0 68.0 48.0 44.0 44.0 46.0 50.0 12.0 ANLI-r2 64.0 56.0 56.0 38.0 36.0 60.0 60.0 48.0 48.0 48.0 42.0 48.0 8.0 ANLI-r3 68.0 56.0 56.0 46.0 46.0 62.0 60.0 46.0 48.0 56.0 58.0 48.0 20.0 MNLI-m 88.0 90.0 88.0 88.0 88.0 86.0 88.0 90.0 94.0 88.0 80.0 96.0 4.0 MNLI-mm 90.0 90.0 94.0 94.0 92.0 94.0 100.0 94.0 98.0 88.0 88.0 80.0 2.0 SNLI 90.0 88.0 88.0 76.0 74.0 90.0 92.0 96.0 96.0 94.0 86.0 80.0 16.0 QNLI 94.0 94.0 30.0 68.0 56.0 74.0 58.0 60.0 62.0 60.0 56.0 56.0 32.0 RTE 88.0 82.0 74.0 78.0 74.0 82.0 76.0 64.0 72.0 64.0 76.0 68.0 36.0 | | Natural language inference | Commonsense | Sentiment | |------------------|------------------------------|---------------|-------------| | ANLI (R1-R3)

fdb20c5e-eda3-427a-9814-eac823469fa8

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## E Details Of Training And Evaluation Datasets 0 56.0 56.0 32.0 RTE 88.0 82.0 74.0 78.0 74.0 82.0 76.0 64.0 72.0 64.0 76.0 68.0 36.0 | | Natural language inference | Commonsense | Sentiment | |------------------|------------------------------|---------------|-------------| | ANLI (R1-R3) | RTE | CoPA | IMDB | | CB | SNLI | HellaSwag | Sent140 | | MNLI | WNLI | PiQA | SST-2 | | QNLI | StoryCloze | Yelp | | | Reading Comp. | | | | | Reading Comp. W/ | | | | |

af0d4334-221a-4885-9c09-6f39904f81c3

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## E Details Of Training And Evaluation Datasets | | | | | Reading Comp. W/ | | | | | Coreference | | | | | commonsense | | | | | BoolQ | RTE | | | | CosmosQA | | | | | CB | SNLI | | | | ReCoRD |

f2db7c9e-3751-4dee-8a99-648f38c9e37e

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## E Details Of Training And Evaluation Datasets | | | CB | SNLI | | | | ReCoRD | | | | | QNLI | | | | information sources. We use the following datasets: (1) ARC, (2) NQ, and (3) TriviaQA. Paraphrase Detection: This task requires models to ascertain whether two sentences convey the same meaning, indicating semantic equivalence. We use the following datasets: (1) MRPC, (2) QQP, and (3) Paws Wiki. Coreference Resolution: Involves identifying instances within a text that refer to the same entity, demonstrating an understanding of textual context. We use the following datasets: (1) DPR and (2) WSC273. Reading comprehension: Assesses the capability to derive answers to questions from a provided text containing relevant information. We use the following datasets: (1) BoolQ, (2) DROP, (3) MultiRC, (4) OBQA, (5) SQuADv1, (6) SQuADv2. Reading Comprehension with Commonsense: Merges traditional reading comprehension skills with commonsense reasoning, requiring understanding beyond the explicit text. We use the following datasets: (1) CosmosQA; (2) ReCoRD. Natural

160fcfde-2760-498a-91d4-9ebc75ec7240

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## E Details Of Training And Evaluation Datasets . Reading comprehension: Assesses the capability to derive answers to questions from a provided text containing relevant information. We use the following datasets: (1) BoolQ, (2) DROP, (3) MultiRC, (4) OBQA, (5) SQuADv1, (6) SQuADv2. Reading Comprehension with Commonsense: Merges traditional reading comprehension skills with commonsense reasoning, requiring understanding beyond the explicit text. We use the following datasets: (1) CosmosQA; (2) ReCoRD. Natural Language Inference: Focuses on deducing the relationship between two sentences, determining if the second sentence logically follows from, contradicts, or is unrelated to the first sentence. We use the following datasets: (1) ANLI, (2) CB; (3) MNLI; (4) QNLI; (5) SNLI; (6) WNLI; (7) RTE.

0db819ce-833c-418b-bc79-362efab93dcd

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## F Implementation Details Of Baseline Methods F.1 Moe Baselines We use E to denote the LoRA expert and R to denote the router. The MoE methods can be expressed | Paraphrase | Struct to text | |------------------|------------------| | ParaCrawl EN/ES | | | MRPC | CommonGen | | QQP | DART | | WMT-16 Tr/En | | | PAWS | E2ENLG | | WMT-16 De/En | | | STS-B | WEBNLG | | WMT-16 Ru/En | | | Closed-book QA | | | WMT-16 Fi/En | | | ARC (easy/chal.) | | | DRP | | | WMT-16 Ro/En | | | NQ | | | WSC273 |

004c5c81-3154-42a9-846e-53db7b047e80

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## F Implementation Details Of Baseline Methods F.1 Moe Baselines | | DRP | | | WMT-16 Ro/En | | | NQ | | | WSC273 | | | TQA | | | WMT-14 En/Fr | | in the following way: $$y=\sum_{i=i}^{k}R(x)_{i}E_{i}(x).\tag{5}$$ We implied two variants of the MoE routing mechanism. (1) **Dense Gating.** Following (Zadour et al., 2023), the router network consists of a dense layer with trainable parameter $W_{g}$, and the gating score could be obtained through a softmax function by: $$s_{i}=R(x)_{i}=softmax(W_{g}^{T}x),\tag{6}$$ (2)**Sparse Gate**. To maintain the sparsity while training, we leverage the Gumbel softmax trick as (Muqeeth et al., 2023; Nie et al., 2021), where the router can be written as: $$\hat{R}(x)_{i}=\frac{(log(R(x)_{i})+g_{i})/\tau}{\sum_{i=1}^{k}exp((log(R(x)_{i})+g_{i})/\tau)}\tag{7}$$ where $g_{i}\sim$ Gumbel$(0,1)$ and

c426bd64-6381-4d31-93c8-ac2bac298711

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## F Implementation Details Of Baseline Methods F.1 Moe Baselines max trick as (Muqeeth et al., 2023; Nie et al., 2021), where the router can be written as: $$\hat{R}(x)_{i}=\frac{(log(R(x)_{i})+g_{i})/\tau}{\sum_{i=1}^{k}exp((log(R(x)_{i})+g_{i})/\tau)}\tag{7}$$ where $g_{i}\sim$ Gumbel$(0,1)$ and $\tau$ is the temperature. Due to MoE not being easily scalable and arbitrarily adding new LORAs, we randomly selected a LoRA as an expert for each task cluster in the experiment and trained the corresponding Router's parameters. We randomly selected 20 samples for each task during training to form a unified dataset for parameter training.

5ddd2931-6b0f-49d7-8647-19e6237356fd

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## F.2 Smear SMEAR (Muqeeth et al., 2023) does not perform routing aggregation on the Adapter output but rather aggregates the Adapter at the parameter level. We adopt the same setting as the MoE methods, and the results could be calculated in the following way: $$\Theta_{S M E A R}=\sum_{i=i}^{k}R(x)_{i}\Theta_{i},\eqno(8)$$ where Θi denote the parameter of the LoRA-i.

72bc24d7-cac9-4350-801b-ac3309b345bd

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## F.3 Adaptersoup AdapterSoup (Chronopoulou et al., 2023), for new downstream tasks, retrieves the parameters that need to be involved in aggregation through sentence bert and performs weight-space averaging on these parameters to adapt to the new domain. We have uniformly retrieved 3 LoRAs for mixed-task to test their capabilities under mixed-task conditions.

a02eaf99-43fd-4c8d-a112-a4bfa14b4ce0

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## F.4 Lorahub LoRAhub (Huang et al., 2023) also aggregates 20 LoRAs randomly for new downstream tasks. In order to learn the weight of LoRA, a black-box optimization method is employed to learn the weight of each LoRA without calculating the gradients of the large model. It performs weighted averaging at the parameter level. Similar to the training process of MoE, we randomly selected 20 samples for each task to form a unified training dataset for black-box optimization.

04dd0099-c7ef-4986-936e-291281917643

# X: (b,l,d), M: (b,p) # A: (p,r,d), B: (p,d,r) # LoRA fusion computation FA=torch.einsum('bp,prd->brd',M,A) FB=torch.einsum('bp,pdr->bdr',M,B) mid=torch.einsum('bld,brd->blr',X,FA) ## G More Related Works Personalized LoRA serving. Sheng et al. (2023) propose S-LoRA to discuss serving thousands of concurrent LoRA. The framework targets scenarios in which multiple tasks must be handled simultaneously without compromising the efficiency of the base models. Wen and Chaudhuri (2023) propose FLoRA, which enables efficient batching of diverse request types in low-rank adaptation (LoRA) of foundation models. These studies discuss how to deploy or train personalized LoRAs. However, these methods can only utilize a single user-specified LoRA during inference, failing to fully leverage the combination of LoRAs from different tasks. Moreover, the primary focus of these discussions is on computational strategies in GPUs and training strategies, which are orthogonal to the routing strategies with which we are concerned.

73c8a7cd-a42e-4ec9-ac72-38c1b8131bc7

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy Gioele Barabucci†,1, Victor Shia2,3, Eugene Chu4, Benjamin Harack3,5, and Nathan Fu3 1University of Cologne, 2Harvey Mudd College, 3The Human Diagnosis Project, 4Kaiser Permanente, 5University of Oxford Background Large language models (LLMs) such as OpenAI's GPT-4 or Google's PaLM 2 are proposed as viable diagnostic support tools or even spoken of as replacements for "curbside consults". However, even LLMs specifically trained on medical topics may lack sufficient diagnostic accuracy for real-life applications. Methods Using collective intelligence methods and a dataset of 200 clinical vignettes of real-life cases, we assessed and compared the accuracy of differential diagnoses obtained by asking individual commercial LLMs (OpenAI GPT-4, Google PaLM 2, Cohere Command, Meta Llama 2) against the accuracy of differential diagnoses synthesized by aggregating responses from combinations of the same LLMs. Results We find that aggregating responses from multiple, various LLMs leads to more accurate differential diagnoses (average accuracy for 3 LLMs: 75.3% ± 1.6pp) compared to the differential diagnoses produced by single LLMs (average accuracy for single LLMs: 59.0% ± 6.1pp). Discussion The use of collective intelligence methods to synthesize differential diagnoses combining the responses of different LLMs achieves two of the necessary steps towards advancing acceptance of LLMs as a diagnostic support tool: (1) demonstrate high diagnostic accuracy and (2) eliminate dependence on a single commercial vendor.

15e64814-b4f6-4123-93f5-c486f9a6bb22

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 1 Background Large language models (LLMs) such as GPT-4 have been shown to be useful as support tools in various healthcare settings such as during tumor boards [1] or as a screening tool to match patient notes to best practice alerts [2]. Their future use and deployment in healthcare is expected to parallel that of other AI tools such as automated electrocardiogram (ECG) anomaly detection, i.e., as support tools that provide insight to human practitioners to better inform their decisions [3]. In particular, there is ongoing research into the application of LLMs as summarization tools for patient and procedure information [4] or as replacement for "curbside consults", especially in situations where colleagues may not be available (e.g., remote locations) or too expensive (e.g., underserved groups) [5, 6]. Nevertheless, the use of LLM-based tools is impaired by their limited acceptance by medical professionals, among other factors [7]. One major factor driving this low rate of acceptance is lack of trust that an LLM can provide correct answers, and additionally, one that avoids so called "hallucinations", i.e., verisimilar but fictional responses. This lack of trust can in turn be traced back to lack of data on the performance (e.g., correctness, accuracy, specificity) of said LLM-based tools. Put bluntly, before trusting them, medical practitioners want to know: "[Are they] good enough?" [5]. Recent studies on the accuracy of LLMs have been shown them to be capable of performing well in certain medical contexts, but results seem to also vary depending on the research study or application, which paints an unclear situation. For instance, Eriksen *et al.* [8] report that GPT-4 scores above 72% of the readers of medical journals in 38 clinical case challenges. On the other hand, Barile et al. [9] report that GPT-4 has a diagnostic error rate of 83% when confronted with 100 pediatric case challenges published on JAMA and NEJM. This study investigates the use of collective intelligence methods to synthesize higher-accuracy differential diagnoses by aggregating differential diagnoses produced by a set of LLMs (even ones with low accuracy) in response to medical questions in the form of case vignettes. Aggregating results from multiple LLMs could be the key to achieving high-accuracy responses (and potentially with fewer implausible or "hallucinated" responses).

de332db1-76f3-4b3e-935f-789d43ddc04f

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 1 Background hand, Barile et al. [9] report that GPT-4 has a diagnostic error rate of 83% when confronted with 100 pediatric case challenges published on JAMA and NEJM. This study investigates the use of collective intelligence methods to synthesize higher-accuracy differential diagnoses by aggregating differential diagnoses produced by a set of LLMs (even ones with low accuracy) in response to medical questions in the form of case vignettes. Aggregating results from multiple LLMs could be the key to achieving high-accuracy responses (and potentially with fewer implausible or "hallucinated" responses). By employing algorithmic methods for knowledge aggregation from research on collective intelligence, it is possible to create a high-quality response to a question by aggregating lower-quality responses from multiple respondents [10]. For instance, past studies in the medical domain show that aggregating as few as three answers from inexperienced respondents led to high diagnostic accuracy (77% accuracy), significantly above the performance of individual human experts (62.5% accuracy) [11].

b05b800d-0d53-4978-a67d-5c57cf8104d6

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 2 Methods This study can be summarized as follows: we sampled 200 clinical vignettes of real-life cases from the Human Diagnosis Project (Human Dx) database, asked various LLMs to provide differential diagnoses for these cases, aggregated their responses using collective intelligence algorithms, and finally compared the accuracy of the individual LLM responses to the accuracy of the aggregated differentials. The prompts and the Python scripts used to run the study are provided in the supplemental material. The case dataset and the LLM responses are available upon request.

232d6a55-ae36-4cad-b8f3-d5622f8772d8

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 2.1 Dataset And Case Selection The data for this study is a set of 200 case vignettes, extracted from the Human Dx database of clinical cases. Human Dx is a multinational online platform in which physicians and medical students solve teaching cases, as well as offer clinical reasoning support to fellow users. The 200 cases have been randomly sampled from the dataset used by Barnett *et al.* [11], restricting the sampling to text-only vignettes. The correct diagnosis of each case (*ground truth*) is known and has been validated by medical experts as part of the Barnett *et al.* [11] study.

4c5b4162-2cac-4bad-8b5a-17bc15ef7d59

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 2.2 Querying Of Llms Four general-purpose LLMs were asked to solve each case by providing a differential with five ranked diagnoses. The four LLMs used in this study are: OpenAI GPT-4, Google PaLM 2 for text (text-bison), Cohere Command, Meta Llama 2 (llama-2-70b-f). All prompts used to query the LLMs follow the same template: "*[CASE TEXT]* What is the differential (list format of common shorthand nonabbreviated diagnoses) for the above case? Respond with ONLY diagnosis names (one per line) up to a max of 5.", where *[CASE TEXT]* is replaced with the textual description of the case vignette. The actual prompts vary slightly between different LLMs because of different query paradigms (e.g., GPT-4 uses a *chat* paradigm while PaLM 2 uses a text generation paradigm). In order to obtain cleaner differential diagnoses for combining in collectives in the scoring process, a round of manual prompt engineering has been carried out using less than 5 Human Dx case vignettes to help with the format and structure of the response. We are highly confident that the case vignettes used in this study are not part of the training corpora of these LLMs. First, because the case vignettes are only available to users logged into the Human Dx application and to select research partners. Second, because there are contractual agreements in place between Human Dx and the providers of the LLMs that forbid the use of data included in prompts as training material (except for Cohere). Finally, the correct diagnoses were never included in any of the prompts to the various LLMs.

db00d4dd-1bc8-49fd-9057-3255e929eef0

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 2.3 Collective Intelligence And Response Aggregation Starting from the differential diagnoses provided by the single LLMs, 11 synthetic differential diagnoses have been generated by aggregating the single differential diagnoses in all possible combinations (6 two-fold combinations, 4 three-fold combinations, 1 four-fold combination). The aggregation method is a frequency-based, 1/r-weighted method similar to those used in other collective intelligence studies focused on diagnostic tasks via differential diagnoses [11, 12]: 1. *Normalization*: All diagnoses in the differentials are normalized by removing common prefixes (e.g., "syndrome", "disorder"), stop words (e.g., "by", "of", "with"), and punctuation signs. In addition, synonyms are merged into preferred terms, following the matching established by Barnett *et al.* [11]. 2. *Extraction of unique diagnoses*: The set of all unique normalized diagnoses present across all differentials is created. 3. 1/r *weighting*: Inside each differential each diagnosis is given an *individual score* calculated as the inverse of the rank r of the diagnosis in the differential (i.e., the first diagnosis is given the score 1/1 = 1, the second 1/2 = 0.5, the third 1/3 = 0.33, etc). 4. *Aggregation*: Each of the unique diagnoses in the set created in step 2 is given a aggregate score calculated by adding all the individual scores of that diagnosis across all differentials. 5. *Synthesis*: A synthetic differential is generated by taking the five unique diagnoses with the highest aggregate score and ranking them by their score in decreasing order.

c4d55222-9dd6-4b53-a828-f2b41a51692a

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 2.4 Accuracy Measure The accuracy of a solver (either an LLM or a group of LLMs) is calculated as the percentage of correctly diagnosed cases among all cases. For this study we consider a case to be correctly diagnosed by a solver if the differential provided by that solver for that case contains the correct diagnosis among the five highest ranked diagnoses. This so-called *TOP-5 matching* mirrors similar correctness measures used in previous studies [11, 12]. Results obtained using TOP-1 or TOP-3 matching are provided in the supplemental material.

a03a4f82-95ed-43df-9f06-1b2406a8f1b6

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 3 Results The main finding of this study is that the accuracy of differential diagnoses created by aggregating differential diagnoses from multiple LLMs using collective intelligence methods (accuracy for 3 LLMs: 75.3% ± 1.6pp) are consistently better than the accuracy of differential diagnoses produced by single LLMs (average accuracy of single LLMs: 59.0% ± 6.1pp). The average accuracy of individual LLMs is 59.0% ± 6.1pp, i.e., on average a LLM produces a ranked differential diagnosis that contains the right diagnosis in 59.0% of the cases. The average accuracy of groups of LLMs increases as the group size grows: the average accuracy for groups of 2 LLMs is 69.1% ± 2.6pp, for 3 LLMs is 75.3% ± 1.6pp, for 4 LLMs is 78.0% ± 0.1pp. Figure 1 shows the individual and average accuracy of the single LLMs and of groups of LLMs. This finding holds true also when the definition of correctly diagnosed case is made stricter by considering only the 3 highest ranked diagnosis in a differential (TOP-3 matching), as illustrated in Figure 2. This trend is also confirmed when the differential diagnosis of the LLM with the highest individual accuracy, GPT-4, is excluded from the experiment (average accuracy of single LLMs: 54.6% ± 6.4pp, of groups of 2 LLMs: 63.5% ± 2.3pp, 3 LLMs: 70.0% ± 0.1pp). The supplemental material provides data on these alternative evaluation methods.

95edf2fa-f2e0-4fd2-8362-59a7dc293365

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion This study demonstrates the feasibility and validity of using collective intelligence methods to combine low-accuracy differentials from multiple LLMs into synthetic high-accuracy differentials. The degree of increase in accuracy achieved by the method employed in this study is in line with similar results in the field of collective intelligence, both in the medical field [11, 12] and outside [10, 13]. The mechanism that allows this increase in accuracy is that the presented aggregation method emphasizes plausible diagnoses (likely to be present in the differential returned by multiple LLMs, and so, bound to have a higher aggregate score), while minimizing the effects of hallucinated diagnoses (likely to be present in only one of the LLMs). This can be seen as an instance of the Anna Karenina principle (good answers are common to many LLMs, bad answers are local to a specific LLM). This principle is exploited by similar techniques | Group size | LLMs in group | Accuracy | Average accuracy | |--------------|-----------------------------------------------------------|------------|--------------------| | 1 | Cohere Command | | | | 39 | . | 5% | | | 59 | .

c2bfe44a-615a-4065-a3b3-a1cf9e34c72d

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion | 5% | | | 59 | . | 0% | | | ± | | | | | 6 | . | 1 | pp | | 1 | Google PaLM 2 | | | | 66 | .

0dfb6821-a91f-4548-be5c-d56255716bcc

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion | | 1 | Google PaLM 2 | | | | 66 | . | 0% | | | 1 | Meta Llama 2 | | | | 58 | . | 5% | | | 1 | OpenAI GPT-4 | | | | 72

5c0406df-7a7e-422a-96d5-66cac2ba76f3

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion | | 1 | OpenAI GPT-4 | | | | 72 | . | 0% | | | 2 | Cohere Command, Meta Llama 2 | | | | 58 | . | 0% | | | 2 | Google PaLM 2, Cohere Command | | | | 64 | .

25cdbe53-2869-464d-b6c5-54c9f38800bd

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion | | | 2 | Google PaLM 2, Cohere Command | | | | 64 | . | 5% | | | 2 | Google PaLM 2, Meta Llama 2 | | | | 68 | . | 0% | | | 2 | OpenAI GPT-4, Cohere Command | | | | 73 | .

b795e46a-90f7-4fec-8e9b-f5f62aa9d5e9

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion | | 2 | OpenAI GPT-4, Cohere Command | | | | 73 | . | 5% | | | 2 | OpenAI GPT-4, Google PaLM 2 | | | | 77 | . | 0% | | | 2 | OpenAI GPT-4, Meta Llama 2 | | | | 73 | .

a764e5c5-1b3f-49b9-826e-69d4b2862807

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion | | 2 | OpenAI GPT-4, Meta Llama 2 | | | | 73 | . | 5% | | | 3 | Google PaLM 2, Cohere Command, Meta Llama 2 | | | | 70 | . | 0% | | | 75 | . | 3% | | | ± |

f499d5d3-1693-4bda-ba68-f00578c54fc9

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion | | 75 | . | 3% | | | ± | | | | | 1 | . | 6 | pp | | 3 | OpenAI GPT-4, Cohere Command, Meta Llama 2 | | | | 75 | . | 5% | | | 3 | OpenAI

43039592-bc2b-4e24-8e59-28b2c747ad73

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion | | 75 | . | 5% | | | 3 | OpenAI GPT-4, Google PaLM 2, Cohere Command | | | | 79 | . | 0% | | | 3 | OpenAI GPT-4, Google PaLM 2, Meta Llama 2 | | | | 77 | . | 0% | | | 4 | Google PaLM 2, Cohere Command, Meta Llama 2, OpenAI GPT-4 |

02c9b99a-0ee2-4141-83a0-c6c1b1895bbb

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion | . | 0% | | | 4 | Google PaLM 2, Cohere Command, Meta Llama 2, OpenAI GPT-4 | | | | 78 | . | 0% | 78 | | ± | | | | | 0 | . | 1 | pp | used in LLM research such as ensemble methods [14, 15] or multi-agent debates [16]. Two factors differentiate the method employed in this study from these techniques: universality and simplicity

632af423-1cb0-4c66-812f-46eac3a81ada

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion | 1 | pp | used in LLM research such as ensemble methods [14, 15] or multi-agent debates [16]. Two factors differentiate the method employed in this study from these techniques: universality and simplicity. First, this method works despite using LLMs with varying querying approaches and technical differences, and can thus be easily extended to work with any combination of LLMs. Second, the simplicity of this method means that not only can it be easily integrated in existing software applications, but could even be performed by medical personnel manually querying separate LLMs and synthesizing the results themselves. The trust of medical practitioners in LLM-based tools could be strengthened by the application of aggregation methods like the one employed in this study. In particular, knowing that multiple sources 69.1% ± 2.6pp contributed would increase clinician confidence in the final differential and lessen the fear of having encountered one of the many mistaken answers or hallucinations that LLMs are known to produce [5]. An additional advantage of the use of knowledge aggregation methods is preventing vendor lockin, removing the need to engage with a single, potentially expensive, or legally problematic LLM vendor in order to obtain high-quality diagnostic differentials. The use of aggregation methods like the one we propose would address these issues by enabling the use of multiple cheaper LLMs, or alternatively, locally-deployed and fine-tuned LLMs. For instance, our results show that the 3-LLM group without GPT-4 (the top-performing LLM) offers a diagnostic accuracy within a couple of percentage points of GPT-4 alone. With a clear baseline on diagnostic accuracy and trust, LLM-based tools can become valuable support instruments that can speed up diagnosis, reduce diagnostic mistakes and costs, and provide additional consulting services in underserved areas. While this study shows that aggregating differentials produced by current LLMs leads to improved diagnostic accuracy, further studies are needed to examine the impact of future LLMs specialized on medical topics or

7eca7d7d-1b2f-4d41-85b5-069e08e51b79

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 4 Discussion instance, our results show that the 3-LLM group without GPT-4 (the top-performing LLM) offers a diagnostic accuracy within a couple of percentage points of GPT-4 alone. With a clear baseline on diagnostic accuracy and trust, LLM-based tools can become valuable support instruments that can speed up diagnosis, reduce diagnostic mistakes and costs, and provide additional consulting services in underserved areas. While this study shows that aggregating differentials produced by current LLMs leads to improved diagnostic accuracy, further studies are needed to examine the impact of future LLMs specialized on medical topics or the use of participatory AI methods (for instance, the synthesis of differentials aggregating responses from both LLMs and human practitioners).

62f2b072-add5-492a-a719-f4edb562f062

# Combining Insights From Multiple Large Language Models Improves Diagnostic Accuracy ## 5 Acknowledgements The authors would like to thank Nikolas Zöller of the Max Plank Institute for Human Development for his valuable and constructive feedback. The authors would also like to thank Irving Lin and Jay Komarneni of the Human Diagnosis Project for their suggested framing and review. This work is supported by the European Union's Horizon Europe Research and Innovation Programme under grant agreement No 101070588 (HACID: Hybrid Human Artificial Collective Intelligence in Open-Ended Domains).

25d193a1-03fa-455c-af72-4c41343119fc

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis Ridwan Taiwo1*, Idris Temitope Bello1,2, Sulemana Fatoama Abdulai1, Abdul-Mugis Yussif1, Babatunde Abiodun Salami3, Abdullahi Saka4, Tarek Zayed1* 1Department of Building and Real Estate, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, China. 2Centre for Advances in Reliability and Safety (CAiRS), Hong Kong Science Park12/F, Building 19W, Pak Shek Kok, NT, Hong Kong, China 3Cardiff School of Management, Cardiff Metropolitan University, Llandaff Campus, Cardiff CF5 2YB, United Kingdom. 4School of Built Environment, Engineering and Computing, Leeds Beckett University, UK Corresponding authors: Ridwan Taiwo (ridwan-a.taiwo@connect.polyu.hk) and Tarek Zayed (tarek.zayed@polyu.edu.hk)

69ce8c27-cf15-4b63-989b-355550555e0f

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Abstract The construction industry is a vital sector of the global economy, but it faces many productivity challenges in various processes, such as design, planning, procurement, inspection, and maintenance. Generative artificial intelligence (AI), which can create novel and realistic data or content, such as text, image, video, or code, based on some input or prior knowledge, offers innovative and disruptive solutions to address these challenges. However, there is a gap in the literature on the current state, opportunities, and challenges of generative AI in the construction industry. This study aims to fill this gap by providing a state-of-the-art analysis of generative AI in construction, with three objectives: (1) to review and categorize the existing and emerging generative AI opportunities and challenges in the construction industry; (2) to propose a framework for construction firms to build customized generative AI solutions using their own data, comprising steps such as data collection, dataset curation, training custom large language model (LLM), model evaluation, and deployment; and (3) to demonstrate the framework via a case study of developing a generative model for querying contract documents. The results show that retrieval augmented generation (RAG) improves the baseline LLM by 5.2, 9.4, and 4.8% in terms of quality, relevance, and reproducibility. This study provides academics and construction professionals with a comprehensive analysis and practical framework to guide the adoption of generative AI techniques to enhance productivity, quality, safety, and sustainability across the construction industry. Keywords: Generative AI; Artificial Intelligence; Large Language Models; LLM; RAG; ChatGPT; Construction Industry

39efda4a-265f-47a7-ab9e-6132fc08aa81

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 1. Introduction The construction industry is one of the most critical sectors of the global economy, accounting for approximately $10 trillion or 13% of global GDP in 2019 [1–3]. It also employs over 220 million workers globally and is often a significant driver of employment [4]. The construction industry encompasses various sub-sectors, such as civil engineering, infrastructure, residential, commercial, and industrial building construction [1–3]. Among these, building construction is the largest and most diverse sub-sector, accounting for about 40% of the global construction output and 50% of the worldwide construction employment [5]. Building construction is a complex and dynamic process that involves multiple stakeholders, such as owners, architects, engineers, contractors, subcontractors, suppliers, and regulators [6,7]. The process requires coordinating and integrating various activities, such as design, planning, scheduling, procurement, fabrication, installation, inspection, and maintenance [8,9]. The process also generates and consumes vast data, such as drawings, specifications, contracts, reports, invoices, and photos[10]. Building construction's quality, efficiency, and sustainability depend primarily on how well these activities and data are managed and utilized. However, the construction industry faces many challenges that hinder its performance and productivity. These challenges include the design, construction, procurement and supply chain, fabrication and installation, and inspection and maintenance of buildings, as shown in Figure 1 [11–13]. Each of these aspects involves complex and dynamic processes that require coordinating and integrating various resources, disciplines, and stakeholders and considering and managing various constraints, uncertainties, and changes. These challenges pose significant difficulties and risks for the construction industry, resulting in low productivity, high cost, long delays, poor quality, and high environmental impact. According to a report by McKinsey, the global construction industry has an average annual productivity growth of only 1%, compared to 2.8% for the total world economy and 3.6% for manufacturing. The report also estimates that the global construction industry could save up to $1.6 trillion per year by improving its productivity to the level of other sectors [1]. Addressing these challenges and improving the performance and productivity of the building construction industry requires innovative and disruptive solutions that can leverage the power of data and technology. One of the most promising and emerging solutions is generative artificial intelligence (AI) [14–16]. Generative AI is a branch of AI that aims to create novel and realistic data or content, such as text, image, video, audio, or code, based on some input or

55e3b4b8-f14e-4368-a574-d4d836acf8fa

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 1. Introduction 3.6% for manufacturing. The report also estimates that the global construction industry could save up to $1.6 trillion per year by improving its productivity to the level of other sectors [1]. Addressing these challenges and improving the performance and productivity of the building construction industry requires innovative and disruptive solutions that can leverage the power of data and technology. One of the most promising and emerging solutions is generative artificial intelligence (AI) [14–16]. Generative AI is a branch of AI that aims to create novel and realistic data or content, such as text, image, video, audio, or code, based on some input or prior knowledge [17]. Generative AI can be seen as the opposite of discriminative AI, which aims to classify or recognize data or content, such as identifying objects in an image or translating text from one language to another. Generative AI can also be seen as a form of creative AI, which aims to produce data or content that is not only realistic but also original, diverse, and expressive [18]. Generative AI also relies on large language models (LLMs), which are neural network models that can generate natural language text based on a given prompt or context. LLMs are trained on massive amounts of text data from various sources, such as books, articles, websites, and social media. They can capture natural language's semantic and syntactic patterns and relationships [19]. Generative AI has been adopted and applied in various disciplines and domains. For instance, in healthcare, generative models have been used to generate synthetic medical images to address data privacy issues and augment datasets to improve disease detection models [20]. Pharmaceutical companies are also exploring generative chemistry to develop new molecular structures for drug discovery [21,22]. Similarly, businesses have leveraged generative design to create new product ideas early in the conceptual process by automating styles and prototyping [23]. In the social sciences, generative AI enhances historical analysis by restoring damaged documents and generating realistic synthetic population data to explore societal issues [24]. Academia is investigating developed personalized study tools and interactive simulated experiences to complement traditional education [25]. Although generative AI applications are still at an infant stage in the construction industry, a few studies exist on their usage for construction-related work [26,27]. As such, [28] reviewed applications of generative AI for developing and enhancing structural designs and how it could help improve accuracy in the design process. [29] presented an overview of the potential applications of Generative Pre-trained Transformer (G

0478c54a-5a7a-437e-a776-c9104cd204ac

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 1. Introduction social sciences, generative AI enhances historical analysis by restoring damaged documents and generating realistic synthetic population data to explore societal issues [24]. Academia is investigating developed personalized study tools and interactive simulated experiences to complement traditional education [25]. Although generative AI applications are still at an infant stage in the construction industry, a few studies exist on their usage for construction-related work [26,27]. As such, [28] reviewed applications of generative AI for developing and enhancing structural designs and how it could help improve accuracy in the design process. [29] presented an overview of the potential applications of Generative Pre-trained Transformer (GPT) models across the lifecycle of a construction project and a case study for material selection. Further, opportunities and a limited number of challenges of adopting generative AI in the construction industry were presented by [30]. Despite the potential and promise of generative AI for the construction industry, there needs to be more systematic and comprehensive literature that reviews and analyzes the current state, opportunities, and challenges of generative AI in this domain. Most existing literature focuses on specific applications or aspects of generative AI (such as GPT models or text-text models) without considering the broader and holistic picture of generative AI in the construction industry. Moreover, there is an urgent need for a more practical and actionable guidance on implementing and deploying generative AI solutions in the construction industry, especially for construction firms that may need more data, expertise, or resources to develop their own generative AI models from scratch. Therefore, this study aims to provide a state-of-the-art analysis of generative AI in the construction industry, with the following objectives: - To provide potential opportunities and challenges of applying generative AI in the construction industry by reviewing and categorizing the existing and emerging generative AI applications and use cases. - To propose an implementation framework for construction firms to build customized generative AI solutions using their data by describing and explaining the key steps and components of the framework. - To demonstrate the proposed framework via a practical use case of developing a tailored generative model for contract documents. The rest of the paper is organized as follows: Section 2 briefly overviews the foundational algorithms and large generative models. Section 3 describes the study's methodology, which is explained in four phases. Section 4 presents the literature review results and expert discussion, including the current applications, opportunities, and challenges of generative AI in the construction industry. Section 5 proposes the framework for building custom LL

b715d1ac-9641-494b-9319-7f5cfc351a17

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 1. Introduction for construction firms to build customized generative AI solutions using their data by describing and explaining the key steps and components of the framework. - To demonstrate the proposed framework via a practical use case of developing a tailored generative model for contract documents. The rest of the paper is organized as follows: Section 2 briefly overviews the foundational algorithms and large generative models. Section 3 describes the study's methodology, which is explained in four phases. Section 4 presents the literature review results and expert discussion, including the current applications, opportunities, and challenges of generative AI in the construction industry. Section 5 proposes the framework for building custom LLM in the construction industry and explains the main steps and components of the framework. Section 6 demonstrates the framework via a case study of developing a generative model for contract documents and shows the results and outcomes of the case study. Section 7 concludes the paper and provides some directions for future research.

562954b8-82ac-4887-947d-3f8ed0da59d0

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2. Generative Ai Generative AI is a branch of AI that aims to create novel and realistic data or content, such as text, image, video, audio, or code, based on some input or prior knowledge [31]. In contrast to discriminative AI [32,33], which aims to categorize or identify existing data, generative AI focuses on creating new, original data or content. For example, while discriminative models may label objects in an image, generative models can produce new images or texts based on specified inputs[24,28]. Generative AI is based on various foundational algorithms, such as generative adversarial networks (GANs), autoregressive models, variational autoencoders (VAEs), transformer models, diffusion models, normalizing flows, and energy-based models. These algorithms use different techniques, such as adversarial learning, probabilistic modeling, attention mechanism, denoising, and density estimation, to learn the underlying distribution or structure of the data or content and generate new samples or variations. Generative AI also relies on large language models (LLMs), which are neural network models that can generate natural language text based on a given prompt or context. LLMs are trained on massive amounts of text data from various sources, such as books, articles, websites, and social media. They can capture natural language's semantic and syntactic patterns and relationships [19,29]. While LLMs have driven advances in text generation, generative AI broadly refers to various techniques for synthesizing novel data or content across modalities. Beyond natural language, generative models can produce images, 3D models, audio, video, and more.

accf8229-2456-4804-8254-6882cd1c9402

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.1. Foundational Algorithms The foundational algorithms of generative AI are the core methods and techniques that enable the generation of data or content. These algorithms can be categorized into different types based on underlying principles and assumptions. Figure 2 provides a schematic representation of these algorithms and their functionalities in generating/encoding, decoding/reconstructing/translating texts, images, and audio. Figure 2: Schematic summarizing the foundational algorithms for generating/encoding, decoding/reconstructing/translating images, text, and audio. GANs (generative adversarial networks), ARs (autoregressive models), VAEs (variational autoencoders), DMs (diffusion models), TMs (transformer models), NFs (normalizing flows), and EBMs (energy-based models).

0332edae-729a-4e86-885c-65b0b642b76c

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.1.1. Generative Adversarial Networks (Gans) GANs are generative AI algorithms that use a game-theoretic approach to generate data or content by competing between a generator and a discriminator. The generator tries to produce realistic and diverse data or content, while the discriminator tries to distinguish between the real and the fake data or content [34]. GANs can generate data or content, such as images, videos, and audio, based on some input or noise. Still, they can suffer from mode collapse, where the generator produces limited variations of data or content. **Table 1** shows common GAN-based models and their functions.

fd65c0e4-cf36-4f05-9bb7-2d19d2f728b4

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 1: Gan-Based Models And Their Functions | Generative | adversarial | Function | |-----------------------------|----------------|---------------------------------------------------------| | networks | | | | [35] | StyleGAN | A model that can generate high-quality and diverse | | images of human faces | | | | [36] | CycleGAN | A model that can translate images from one domain to | | another without paired data | | | | [37] | BigGAN | A model that can generate large-scale and high-fidelity | | images of various

b8641503-e047-4a1f-8867-bf27adc4e1b0

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 1: Gan-Based Models And Their Functions | | | [37] | BigGAN | A model that can generate large-scale and high-fidelity | | images of various classes | | |

d2d41a28-0515-47fb-9222-b5b619d56f55

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.1.2. Autoregressive Models (Ams) AMs are a generative AI algorithm that uses a sequential approach to generate data or content by predicting the next element based on the previous elements. The model learns the conditional probability distribution of the data or content and its samples to generate new data or content. Autoregressive models can generate sequential and structured data or content, such as text, audio, and code, based on some input or context. A limitation is exposure bias - during training, AMs only see ground truth sequences, not their predictions. So errors can accumulate at inference [38]. **Table 2** presents examples of AM-based models and their key capabilities.

6703fa7d-7060-4225-88a1-deaae0315335

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 2: Autoregressive Models And Their Functions Function Ref Autoregressive models PixelRNN A model that can generate realistic images pixel-by-pixel [39] | [40] | WaveNet | A model that can generate realistic speech and music | |-----------|------------|-------------------------------------------------------------| | waveforms | | | | [41] | GPT-3 | A model that can generate natural language text for various | | tasks | | |

8a3826c2-7316-461d-a793-e7d5afe7fd54

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.1.3. Variational Autoencoders (Vaes) VAEs rely on a probabilistic approach for generating novel data or content. These models map inputs into a latent space and then sample from a prior distribution. VAEs can produce varied outputs that resemble the original data by learning to minimize the reconstruction error and divergence between prior and posterior distributions. Images, videos, audio, and other modalities can be generated by sampling from the latent space based on random noise or conditional inputs [42]. However, VAEs face the challenge of posterior collapse, where the latent space fails to capture meaningful variational information to condition the generations. Some VAE-based models are presented in **Table 3**. Function Ref Variational autoencoders [42] VAE A model that can learn a latent representation of data and generate new data [43] CVAE A model that can learn a conditional latent representation of data and generate new data [44] VQ-VAE A model that can learn a discrete latent representation of data and generate new data

d7c213a2-1df3-4fb2-a912-c956286a85a6

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.1.4. Transformer Models (Tms) Transformers represent a paradigm shift in generative modeling underpinned by attention mechanisms. Rather than sequence alignment, self-attention layers allow modeling long-range dependencies in data by learning correlations between input and output vectors. This gives transformers a global receptive field for generation compared to RNNs' localized windows [29,45]. Beyond sequences like text and audio, transformers can generate graphs, 3D points, and other structural data using self-attention. However, challenges remain. Without explicit alignment, representing order and continuity is difficult. Catastrophic forgetting of rare sequences also occurs as new training data overrides previously learned patterns. Restricting self-attention to local neighborhoods may improve stability [46]. **Table 4** presents common transformer-based models and their functions. Function Ref. Transformer models [47] GPT-4 A model that can perform sequence-to-sequence tasks using attention mechanisms [48] BERT A model that can learn bidirectional representations of natural language for various tasks [49] ViT A model that can learn visual representations of images using transformers

fae431fc-6a31-43c8-89e8-66ae9905a7b1

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.1.5. Diffusion Models (Dms) Diffusion models generate data through a two-step denoising process. First, the algorithm adds isotropic Gaussian noise to the real data over repeated diffusion steps. This gradually destroys structure while maintaining dimensionality. Next, the process is reversed by a neural network that removes noise step-by-step until pristine samples emerge. Unlike autoregressive methods, diffusion models generate entire data instances simultaneously rather than sequentially [50]. This provides inherent parallelism. However, determining optimal diffusion and denoising schedules remains challenging. Models must also undo all noise perfectly or risk compounding errors. Slow convergence arises from the numerous forward and reverse passes required. Recent innovations like denoising score matching have accelerated diffusion model training [51]. Diffusion-based models are presented in **Table 5**. Function Ref. Diffusion models [52] DDPM A model that can generate realistic images by reversing a diffusion process [53] NCSN A model that can learn the score function of data distribution using noise-conditional score networks [54] DALL-E 2 A model that can generate images from text captions using the diffusion technique

b2a8270c-591e-4476-a910-9dc18633f74d

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.1.6. Normalizing Flows (Nfs) NFs offer an invertible approach to density estimation for generative modeling. Applying a series of bijective mappings can transform data into a latent space where the prior distribution is known [55]. This allows exact likelihood calculation. The change-of-variables formula relates probabilities between spaces through the Jacobian. Normalizing flows learn mappings that maximize the likelihood of the training data under the latent space prior. Sampling new points is achieved by testing the latent prior and reversing the flow. Normalizing flows shine where data lies near a lower-dimensional manifold, not the entire ambient space. However, computational complexity rises linearly with depth due to repeated Jacobian determinants. Trade-offs exist between modeling power and efficiency. Recent work has enhanced flows with concepts like sparsity, conditional inputs, and new invertible layers. Normalizing flows remain promising for manifold learning-based generation [55]. **Table 6** outlines examples of NF-based models and their key capabilities. Function Ref. Normalizing flows [56] NICE A model that can learn a bijective mapping between data and noise using additive coupling layers [55] RealNVP A model that can learn a bijective mapping between data and noise using affine coupling layers [57] Glow A model that can learn a bijective mapping between data and noise using invertible 1x1 convolutions and actnorm layers

9e12bd13-9844-4d09-a16f-981cd3acff64

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.1.7. Energy-Based Models (Ebms) EBMs take a thermodynamics approach to generative modeling by formulating data density as Boltzmann distribution. Lower energies indicate higher probability density. The architecture consists of an energy function assigning scalar values to each data configuration and a sampling procedure to generate new data. Typically, the energy function is represented by a deep neural network like a convolutional autoencoder. The network is trained to assign low energies to observed training examples and higher energies elsewhere. Sampling generates new data by initializing random noise and descending via gradient descent until reaching local energy minima. A significant challenge is mode collapse, where sampling needs to cover the full diversity of densities. Insufficient capacity in the energy network also hinders quality. Modern advances integrate EBMs with MCMC sampling and discriminator networks to improve coverage and sample quality [58]. The physics-inspired energy framework provides a unique generative modeling perspective differing from prevailing probabilistic approaches. **Table 7** shows common energy-based models with their capabilities. Function Ref. Energy-based models [59] RBM A model that can learn a joint distribution of data and hidden variables using a bipartite graph [60] Hopfield network A model that can store and retrieve patterns using a recurrent network [58] EBGAN A model that can generate realistic images using an energy-based discriminator

b24ca6ac-1312-45d5-a779-9f18b884612e

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.2. Large Generative Models Large generative models (LGMs) refer to massive neural networks trained on huge datasets that can generate text, images, audio, video, or code. By exposing the model to large corpora of varied content during pretraining, LGMs learn general representations that empower them to generate original outputs within specific modalities. These models are powered by the foundational algorithms discussed in the previous section. For example, models like GPT-4 and Codex can produce coherent text and functional code based on prompts, respectively [27,29]. DALL-E 3 and Imagen can generate photorealistic images from text descriptions [51,61]. Models trained on audio can synthesize natural human speech or music. LGMs can also combine modalities to enable multimodal generation, such as illustrating input stories with suitable images or accompanying lyrics with fitting music compositions. The defining feature of LGMs is their massive scale in model size, computational requirements, and training data volume, which enables versatile and creative generation spanning from text to images to video to code and in an integrated multimodal fashion. This makes them a multipurpose generative toolbox powered by pretraining on diverse big data. Table 9 lists LGMs released in recent years, their developer, training parameters, release year, and accessibility. | Access | Reference | Release | |-----------------------------|-----------------|------------| | year | | | | LGM | Developer | Training | | parameter | | | | (Billion) | | | | G

bd93d727-99b6-464e-8c8a-2b3ba72f6c35

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.2. Large Generative Models | Developer | Training | | parameter | | | | (Billion) | | | | GPT-4 | OpenAI | 1000 | | Gemini Pro | Google DeepMind | 17 | | Llama 2 | Meta | 2000 | | PaLM | Google | 540 | | Claude | Anthropic | 12 | | DALLE-3 | OpenAI | - | | SDXL | Stability AI | 2.6 | | DALLE-2 | OpenAI | 3.5 | | Dreamfusion Google |

9aba98e3-803e-4962-87e2-d4fc84306bb3

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.2. Large Generative Models | - | | SDXL | Stability AI | 2.6 | | DALLE-2 | OpenAI | 3.5 | | Dreamfusion Google | - | 2021 | | Flamingo | Google DeepMind | 3.2 | | Phenaki | Google | 1.8 | | Codex | OpenAI | 100 | | Galactica | Meta | 120 | | AudioLM | Google | 0.6 | | AlphaTensor DeepMind | 0.5 | 2021 | | DALL-E | OpenAI | 12 | | CLIP | OpenAI | 63

20a58e5b-111b-424d-af26-bf9454cbd025

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.2. Large Generative Models DeepMind | 0.5 | 2021 | | DALL-E | OpenAI | 12 | | CLIP | OpenAI | 63 | | BART | Facebook | 0.4 | | T5 | Google | 11 | | BERT | Google | 0.34 | | GPT-3.5 | OpenAI | 175 | | GPT-2 | OpenAI | 1.5 | | | | | | XLNet Google | [74] | | |

b334bb72-47c8-45dc-bc76-6c7f3b9479e1

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 2.2. Large Generative Models | | | | | XLNet Google | [74] | | | | | | | 0.34 | 2019 | | | | | | | Open source | | | | | | |

de468eac-2951-415d-a113-7bda0ce8ac21

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 3. Methodology A four-phase approach is adopted to achieve the objectives of this study. Figure 1 visualizes these phases, including systematic literature review and retrieval, expert discussion and review, a framework for developing custom LGM in the construction industry, and a case study. I. Phase 1 - Systematic literature retrieval and review: The first step in this phase involves selecting appropriate databases for the literature search. Scopus, Web of Science, and ScienceDirect were chosen due to their broad coverage and rigorous indexing of peer-reviewed publications [82,83]. Keyword identification was conducted iteratively to capture relevant studies at the intersection of generative AI and the construction industry. The final search string consisted of ["Construction industry" OR "architecture engineering and construction industry" OR "AEC industry" OR "AECO industry"] AND ["Generative AI" OR "GenAI" OR "GENAI" OR "Bard" OR "Gemini" OR "GPT" OR "GPT-1" OR "GPT-2" OR "GPT-3" OR "InstructGPT" OR "ChatGPT" OR "Transformer" OR "GPT-4" OR "Llama" OR "LamDA"]. This search returned 79 initial results. We narrowed the search results to 10 potentially relevant studies based on the title and abstract screening. An in-depth review found that only four (4) were original research articles, with two review articles, and the full text of the rest was unavailable or written in languages other than English. Snowball searching expanded the final pool to six (6) peer-reviewed papers at the intersection of generative AI and construction. II. Phase 2 - Expert discussion and review: The limited literature identified in Phase 1 highlighted the need to supplement with expert perspectives, given the nascent state of generative AI adoption in construction. To elicit diverse insights, 15 experts with backgrounds spanning AI research and construction industry practice were identified. Invitations were sent to participate in the study, with 11 experts accepting for a 73% response rate. This panel encompassed university professors in AI and construction engineering, technology directors from major construction firms, and founders of AI startups targeting the architecture, engineering, and construction (AEC) industry. A modified Delphi survey was conducted with the panel to identify opportunities and challenges of applying generative AI in the construction industry. Thematic analysis was then used to extract common themes

b9fe28d6-fb41-44ec-a7db-4239b83e88bb

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 3. Methodology expert perspectives, given the nascent state of generative AI adoption in construction. To elicit diverse insights, 15 experts with backgrounds spanning AI research and construction industry practice were identified. Invitations were sent to participate in the study, with 11 experts accepting for a 73% response rate. This panel encompassed university professors in AI and construction engineering, technology directors from major construction firms, and founders of AI startups targeting the architecture, engineering, and construction (AEC) industry. A modified Delphi survey was conducted with the panel to identify opportunities and challenges of applying generative AI in the construction industry. Thematic analysis was then used to extract common themes from the qualitative responses. This involved codifying the experts' opinions and aggregating them into categories through an iterative process. The goal was to determine areas of consensus as well as unique perspectives. III. Phase 3 - Framework for developing custom LGM in the construction industry: This phase involved synthesizing the literature and expert findings into a methodology construction firms can follow to build custom generative AI solutions using their proprietary data. The framework encompasses construction data collection, dataset curation, training the custom LGM, evaluation, and deployment steps. IV. Phase 4 - Case study: A case study was conducted using generative AI for querying contract documents to demonstrate practical application. The first step involved the selection of the base LLM architecture. OpenAI's GPT-4 model was chosen as the base model due to its state-of-the-art natural language generation capabilities. A retrievalaugmented generation (RAG) system was implemented to improve the base LLM further. This mitigated hallucinated text by grounding outputs in relevant dataset examples. LangChain Library was employed for the development [84]. The performance of the customized LLM was evaluated, and a graphical user interface was developed using Streamlit [85]. This interactive web application enabled testing of the customized generative AI model through prompts.

f825b4e7-b353-42e0-96a7-b343e663668e

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4. Results 4.1. Current Applications This section presents the application of generative AI in the form of LGMs in the construction industry. Based on the systematic review, only six peer-reviewed articles exploring the uses of LLMs in construction were identified. No articles relating to other LGMs, such as large image and video models, were found. The six articles are summarized in **Table 10**, including their objective, methods, and contributions. The reviewed studies demonstrate emerging applications of LLMs, such as GPTs and BERT-based models for construction tasks, including virtual assistance, sequence planning, schedule generation, hazard recognition, risk assessment, and project planning [86–88]. The contributions highlight the potential for LLMs to enhance productivity, accuracy, and automation in areas like information retrieval, education/training, and documentation review. However, the limited number of studies indicates that the adoption of modern generative AI in construction is still in the very early stages. Significant research is needed to develop customized LLMs for the industry and validate their capabilities on realworld problems.

547493be-71c2-4790-bd84-f56bdb165171

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | Reference Objective | Methods | |---------------------------------------------------------------------------------|---------------------------------------| | [89] | Development of a dynamic prompt- | | based virtual assistant framework for | | | BIM information search | | | The framework's application | | | improves information search

8aae0f2d-a83f-474e-b5d2-58161cd23d90

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | The framework's application | | | improves information search | | | speed, accuracy, and user | | | experience. | | | The framework integrates BIM and GPT technologies for an NL-based | | | interface.

a747d10e-9cd6-48db-ae13-fb48cdfce7bd

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | The framework integrates BIM and GPT technologies for an NL-based | | | interface. | | | | | | Dynamic prompt-based process interprets NL queries, retrieves information, | | | and delivers responses. | | | [90]

15bcf052-fe9a-4105-acca-6d81b99eb0c1

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | and delivers responses. | | | [90] | Development of RobotGPT for | | automated sequence planning in | | | robotic assembly for construction | | | tasks. |

2d00ae54-e537-4499-9a4f-30e46cd4129e

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | tasks. | | | RoboGPT-driven robots can | | | handle complex construction | | | operations and adapt to | | | changes on the fly. |

336a16ed-8d1b-4b1d-b98a-192367320a99

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | | changes on the fly. | | | RoboGPT is a system that uses ChatGPT for automated sequence planning in | | | robot-based construction assembly. | | | | | | The experimental evaluation included two case studies and 80 trials involving | | | real construction tasks.

ae990e5f-6954-48aa-9efd-afaa4b31bee3

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | | The experimental evaluation included two case studies and 80 trials involving | | | real construction tasks. | | | [91] | Generation of a construction schedule | | for a project | | | The use of LLM to enhance | | |

25ed3529-daf0-4401-8a85-003ce4920a8e

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | The use of LLM to enhance | | | construction schedules | | | workflow. | | | ChatGPT is employed to generate a construction schedule for a simple project. | | | |

76b0f6c4-a2e7-47b5-8f2d-3c4ffb128ed5

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | | | | A survey was conducted to evaluate output quality and participants' experience. | | | | | | Parameters used to evaluate results include accuracy, efficiency, clarity, | | | coherence, reliability, relevance, consistency, scalability, and adaptability. | | | [92]

12da273f-7ee1-4bd5-b87f-f0289a1deaa5

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry , scalability, and adaptability. | | | [92] | The use of ChatGPT for improving | | hazard recognition on construction site | | | The investigation involved 42 students in a construction program. | | | | | | Pre- and post-intervention hazard recognition abilities were measured. | | | The potential of employing

9ab76757-43ca-4e81-a748-932f11cd006d

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | Pre- and post-intervention hazard recognition abilities were measured. | | | The potential of employing | | | ChatGPT for safety education | | | and training. | | | [93] | Automated classification of | | contractual risk clauses

138a391b-226a-4de9-a100-72bda9f62288

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | [93] | Automated classification of | | contractual risk clauses | | | The model improves the | | | construction specification | | | review process and risk |

e54aa3fe-5830-4e45-8c6c-d34d22b44de2

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | review process and risk | | | management. | | | The BERT method is used for clause classification in construction | | | specifications. | | | |

f26321d6-be67-41f6-a223-b5603c47d24b

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | | | | | Seven risk categories were identified: payment, temporal, procedure, safety, | | | role and responsibility, definition, and reference. | | | [94] | Automatic matching of look-ahead | | planning tasks to master scheduled | | | activities

a768ca7a-0fc8-4998-af7d-8b4d3311935f

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry of look-ahead | | planning tasks to master scheduled | | | activities | | | Both location-based and distance-based matching followed were employed. | | | | | | GPT-2 was used for final matching. | | | Auto-alignment of long-term

948e6611-97f3-4f0c-bdf3-d17036c9731d

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## Table 9: Summary Of Current Applications Of Llm In The Construction Industry | | GPT-2 was used for final matching. | | | Auto-alignment of long-term | | | and short-term plans in | | | construction projects | |

ea626e5b-c6b8-42c8-87fd-e5ca2e766c0d

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2. Opportunities The discussions conducted with experts revealed numerous potential opportunities to deploy generative AI across construction, categorized by input-output capabilities. Sections 4.2.1 through 4.2.9 extensively examine applications that generate text, images, and video. While generative models can also synthesize audio, experts advised that video generation can serve dual visual and auditory content purposes as construction relies heavily on visual data like drawings, photos, animations, and written and verbal communications, generative modes spanning text, images, and video were seen as most directly relevant. Despite significant progress, there are still limitations, particularly when generating complex images and videos. Table 11 summarizes leading generative models for different data types. LLMs like GPT-4 and Gemini Pro demonstrate proficiency in text synthesis [45,95]. DALL-E 3 produces images from text captions [51]. Video generation models like CogVideo, Lumiere, and Stable Diffusion show promise but are still being refined [96,97]. Although there are shortcomings, the pace of progress makes generative AI a promising technology for transforming the construction industry. If trained on sufficient domain data, Text generation achieves high coherence and accuracy. Photorealistic image synthesis provides value in design and documentation use cases. Video capabilities lag but rapidly improve through advances like higher resolution GANs [97,98]. Input-output type Model Developer Reference [45,95] Text to text GPT-3, GPT-4, Gemini Pro OpenAI, Google's DeepMind Text to image DALL-E 3 OpenAI [51] [96,97] Text to video CogVideo, Lumiere Nightmareai, Google Research [45,95] Image to text GPT-4, Gemini Pro OpenAI, Google's DeepMind [99] Image to image Pix2Pix Berkeley AI Research Stability AI [50] Image to video Stable Video Diffusion Video to text VideoCoCa Google Research [100] Video to image - - [97,98] Video to video Lumiere, Gen-2 Google Research, Runway

f7710d4b-a6ed-4bb2-9f1d-9fc864117de5

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.1. Text To Text Generative AI revolutionizes the construction industry by converting textual data into advanced textual outputs, assisting in many tasks in the construction project phases. **Table 12** provides a detailed overview of the potential applications of text generation in the construction industry, categorized using the different project phases. In pre-construction, it can help generate feasibility study summaries, ensure regulatory compliance, and automate proposal/bid drafting [101]. Drafting daily progress reports, specifications, task instructions, and other documents can be automated during construction (see **Figure 4**). Post-construction opportunities include creating inspection reports, punch lists, operation and maintenance manuals, reviewing warranty/compliance letters, and translating documents. Other cross-cutting text applications are information retrieval through natural language queries and translation into multiple languages [16]. With proper training in technical corpora, they can translate industry insights directly into clear, accurate documents without tedious hands-on work. Realizing this potential requires careful, prompt engineering and alignment with construction linguistic patterns and technical jargon. Potential opportunity Description Project phase Pre-construction Generation of the feasibility report summary Summarize extensive feasibility reports and extract key insights and recommendations for informed decision-making during the project initiation. Pre-construction Documentation of regulatory compliance Leverage generative AI to assist in creating documents that ensure compliance with regulatory requirements, a crucial task in the pre-construction planning phase. Pre-construction Preparation of proposal/bid Apply generative AI to assist in the preparation of proposals and bids by automatically generating well-structured and persuasive text content. Construction Generation of daily progress report Create a daily progress report template summarizing on-site activities and achievements during construction. Construction Refinement of construction specifications Utilize generative AI to refine and enhance construction specifications, ensuring clarity and accuracy in the documentation of materials, methods, and standards. Construction Task Assignment and Communication Facilitate task assignment and communication by automatically generating clear and detailed instructions for construction teams through generative AI. Post-construction Summarization of as-built documents Summarize the extensive as-built documentation, providing a condensed overview of the final constructed project for post-construction analysis. Post-construction Generation of facility maintenance manual Automate the generation of comprehensive facility maintenance manuals based on the final as-built documentation. Post-construction Review of warranty and compliance document

dbeb3c17-e0bf-4fff-a394-5fd39b87dbdf

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.1. Text To Text clarity and accuracy in the documentation of materials, methods, and standards. Construction Task Assignment and Communication Facilitate task assignment and communication by automatically generating clear and detailed instructions for construction teams through generative AI. Post-construction Summarization of as-built documents Summarize the extensive as-built documentation, providing a condensed overview of the final constructed project for post-construction analysis. Post-construction Generation of facility maintenance manual Automate the generation of comprehensive facility maintenance manuals based on the final as-built documentation. Post-construction Review of warranty and compliance document Utilize generative AI to review warranty and compliance documents, summarizing critical information and ensuring adherence to post-construction requirements. All Language translation and localization Translate text content between different languages, aiding in global collaboration and communication. All Information retrieval and knowledge discovery Enhance contextual search capabilities by using generative AI to understand and respond to natural language queries, improving the accuracy of information retrieval.

c020480c-f960-43f3-b741-fd83894fd554

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.2. Text To Image AI's text-to-image conversion provides innovative possibilities in the construction field, including the ability to visualize pre-construction architectural ideas, assist in making real-time construction choices, and enhance marketing materials once construction is completed. The potential opportunities for generating images via text prompting are shown in **Table 13**. Generating images from text has broad applicability in construction projects. Pre-construction applications include creating visualizations from site descriptions for selection and planning [102]. Text prompts can also render architectural concepts and project models (**Figure 5**). During construction, progress visualization, equipment layouts, and safety illustrations can be automated from textual inputs. **Figure 6** shows a visualization of construction progress through different stages of execution. Post-construction use cases involve as-built visualization, usage guidelines, and renovation proposals. With appropriate training, models like DALL-E can translate construction domain language into detailed visuals through well-prompted texts. This technology allows people without expertise to readily obtain visual depictions by articulating what they wish to see in plain language. Automating this linkage between vision and language can make project information more accessible while freeing worker time [103]. | Potential opportunity | Description | Project phase | |-----------------------------------------------------------------------------------------|-------------------------------------|--------------------------------------------------------------------------------------------| | Pre-construction | Site

86fc93c3-d986-4e50-ab84-7b58a8ffe866

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.2. Text To Image | |-----------------------------------------------------------------------------------------|-------------------------------------|--------------------------------------------------------------------------------------------| | Pre-construction | Site visualization and selection | Create visual representations of potential construction sites based on text descriptions, | | aiding the decision-making process during site selection. | | | | Pre-construction | Architectural concept rendering | Transform textual architectural concepts into visual renderings, providing stakeholders | | with a clear preview of the proposed designs. |

04a39efb-e407-492d-bf3c-b9143c954203

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.2. Text To Image | Architectural concept rendering | Transform textual architectural concepts into visual renderings, providing stakeholders | | with a clear preview of the proposed designs. | | | | Pre-construction | Interactive project models | Utilize generative AI to convert project descriptions into interactive 3D models, allowing | | stakeholders to explore and engage with the project before construction begins. | | | | Construction

0a6c3f27-35e8-41f7-ab15-9eb49203cb72

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.2. Text To Image | | | Construction | Construction progress visualization | Implement generative AI to generate visual representations of construction progress | | based on textual updates, providing stakeholders with a visual timeline of the project. | | | | Construction | Material and equipment layouts | Through generative AI, create visual layouts of materials and equipment based on | | textual descriptions, optimizing their placement on the construction site. |

bfde6f57-ac69-428c-94b8-ab3cb9339243

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.2. Text To Image | Material and equipment layouts | Through generative AI, create visual layouts of materials and equipment based on | | textual descriptions, optimizing their placement on the construction site. | | | | Construction | Safety procedure illustrations | Apply generative AI to convert text-based safety procedures into visual illustrations, | | enhancing comprehension and adherence to safety protocols on the construction site. | | | | Post-construction

51ac8ce2-e149-41c3-9117-8c8023908027

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.2. Text To Image | | | Post-construction | As-Built visualization | Transform the as-built documentation into visual representations, aiding in the | | visualization and analysis of the final construction. | | | | Post-construction | Facility usage guidelines | Create visual guidelines for facility usage based on textual documentation, ensuring clear | | communication of post-const

749265ca-80c0-466d-a99a-381d0ddf5eef

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.2. Text To Image | | Post-construction | Facility usage guidelines | Create visual guidelines for facility usage based on textual documentation, ensuring clear | | communication of post-construction guidelines. | | | | Post-construction | Renovation proposal visualizations | Generate visual representations of proposed renovations, aiding decision-making during | | the post-construction phase. | |

767b3286-bb1c-4122-b2a2-c16b633beb1c

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.2. Text To Image | Generate visual representations of proposed renovations, aiding decision-making during | | the post-construction phase. | | | | All | Project timeline infographics | Convert textual project timelines into visual infographics, providing an easily | | understandable overview for all project phases. | |

388ed7a8-77e0-4eee-9976-1b3df780e570

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.2. Text To Image | | | | All | Project dashboard visuals | Generate visual representations for project dashboards based on textual data, offering | | stakeholders an intuitive and informative overview of project metrics. | | | | |

68777e4b-9670-42a0-b8f2-766e93222ae9

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.2. Text To Image | | | | |

e5241d46-95c5-468e-b066-54858fa07133

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.3. Text To Video The utilization of generative AI to transform textual information into dynamic video content offers numerous benefits. **Table 14** summarizes key opportunities for text-to-video generation in construction based on the expert discussion. During the pre-construction phase, introductory site exploration videos and animated project concept videos could be synthesized from text to aid scope planning and stakeholder intelligence. During construction, step-by-step equipment operation tutorials and safety training animations could be generated from manuals and textual hazard narrations, respectively [104]. Progress update videos compiled from schedules and logs would help keep stakeholders informed with matching visual updates. Post-construction use cases include creating instructional facility usage videos from the documentation. With appropriate training data, text-to-video models can translate construction domain language into vivid animations and live footage [96,97]. Rather than relying solely on static diagrams and dense text, bringing instructions and processes to life through AI-generated videos makes project information more engaging. Dynamic video tutorials personalized via text to each situation may enhance comprehension and learning for safety training and equipment operation. Automating the linkage between textual descriptions and video footage also frees workers time spent manually storyboarding and editing visualizations. As text-to-video generation techniques continue advancing in resolution and realism, the applications across the construction project lifecycle will expand. | Potential opportunity | Description | Project phase | |----------------------------------------------------|-----------------------------|--------------------------------------------------| | Pre-construction | Site introduction videos | Create introductory videos for potential | | construction sites, providing stakeholders with | |

b994e84e-4668-43c1-98e1-c265b980fed8

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.3. Text To Video | Site introduction videos | Create introductory videos for potential | | construction sites, providing stakeholders with | | | | visual overviews based on textual descriptions. | | | | Pre-construction | Project concept animation | Transform textual project concepts into animated | | videos, offering stakeholders a dynamic | | | | visualization of the proposed construction. | | | | Construction

ede937fb-6dce-4822-8548-9e492b9d034e

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.3. Text To Video the proposed construction. | | | | Construction | Equipment operation guides | Generative AI can automatically create step-by- | | step video tutorials demonstrating equipment use | | | | from the text and diagrams in instruction manuals. | | | | Construction | Safety procedure animations | Safety managers could compose comprehensive | | narrations of hazards and precautions. Generative | | | | AI can synthesize engaging video footage |

d6a93f0d-1d1c-4411-85f9-fef001238d9f

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.3. Text To Video | | narrations of hazards and precautions. Generative | | | | AI can synthesize engaging video footage | | | | matching the narration to create safety training | | | | materials. | | | | Construction | Progress update videos | Generate progress videos automatically using | | generative AI using progress reports, schedules, |

178091d9-05cc-4c7f-a30f-c94d7d25d8d4

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.3. Text To Video | | Construction | Progress update videos | Generate progress videos automatically using | | generative AI using progress reports, schedules, | | | | logs, and notes. | | | | Post Construction | Facility usage instruction | | | videos | |

b124d206-de19-418b-ac0e-f8101a76028c

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.3. Text To Video | | videos | | | | Generate instructional videos based on textual | | | | documentation for facility usage, ensuring clear | | | | communication of post-construction guidelines. | | | | Post Construction | Building update videos | Produce AI-generated videos summarizing facility | | modifications, upgrades, and status changes over |

133ee94e-4d0b-4cfd-895d-4c488c63a8b4

# Generative Ai In The Construction Industry: A State-Of-The-Art Analysis ## 4.2.3. Text To Video | | Post Construction | Building update videos | Produce AI-generated videos summarizing facility | | modifications, upgrades, and status changes over | | | | time from text-based building logs for | | | | stakeholders. | | | | All | Project journey montage | Implement generative AI to compile a video | | montage showcasing the entire project journey, |