ResCity-LLM Pipeline Turns Urban Resilience Knowledge Into Specialized AI Models

A team of urban resilience and earthquake engineering researchers has unveiled a complete pipeline for building large language models that can absorb the sprawling, fragmented knowledge of resilient cities and then deploy it across professional disaster-prevention tasks. The framework, called ResCity-LLM, links together four stages that had previously existed in isolation: assembling a domain corpus, generating training data automatically, adapting open-source models through a two-stage training strategy, and evaluating the results on a purpose-built benchmark. When the strongest adapted model was tested against two much larger general-purpose systems, it came out ahead on the benchmark’s overall weighted score, offering early evidence that compact, locally deployable models can compete with cloud giants in a specialized engineering domain.

The motivation stems from an uncomfortable mismatch between what urban resilience research demands and what general-purpose language models are trained to do. Urban resilience draws on engineering, environment, planning, ecology, economics, sociology and governance simultaneously, with relevant knowledge scattered across academic papers, technical standards, planning documents, policy texts and case studies. General-purpose models such as ChatGPT, DeepSeek and Claude are trained overwhelmingly on general text, so they struggle to cover the complex knowledge structures, specialized semantics and typical task scenarios of this interdisciplinary field. Unlike medicine or law, where domain-specific language model research is relatively mature, resilience-oriented work had remained scattered and exploratory, lacking a systematic framework that spans corpus organization, training data construction, model training and specialized benchmark evaluation.

The first ingredient is a data foundation the researchers call ResCity-Corpus. Using keywords related to urban and resilient cities in both English and Chinese, the team harvested academic literature from Elsevier, Nature, Springer, arXiv and the China National Knowledge Infrastructure, crawled news reports from mainstream media platforms, and collected professional monographs together with national, industry and local technical standards covering disaster prevention, urban lifeline systems and related fields. The final processed corpus for continued pre-training contained 60,059 text blocks amounting to 96.80 million tokens, split roughly 69 percent English and 31 percent Chinese, and organized into seven thematic categories ranging from hazard and climate science to governance and public policy.

Raw corpora of this kind are notoriously messy, so the researchers designed what they call a multi-agent collaborative dual-stage data generation pipeline. In the first stage, aimed at continued pre-training, a clean-agent parses heterogeneous documents and strips out noise, a chunk-agent segments long texts into semantically coherent units sized by tokenizer length with overlapping tails to preserve context, a quality-agent powered by a commercial language model scores each chunk and flags low-quality or incomplete samples, and a repair-agent performs minimal fixes to broken sentences and grammatical errors without altering the underlying knowledge content. This produces continuous text suitable for autoregressive pre-training rather than mechanically cut fragments, with every chunk carrying metadata, token statistics and quality flags.

The second stage converts the same corpus into supervised fine-tuning data by generating instruction-and-response pairs, and here the pipeline splits into two complementary branches. Knowledge-oriented data consists of question-answer pairs centered on concepts, mechanisms and principles, built by an evidence-agent that extracts and scores semantic evidence spans. Task-oriented data instead introduces realistic problem contexts such as urban planning, risk assessment and resilience quantification, built by a knowledge-agent that extracts reusable knowledge units which a generation-agent then wraps into scenario-driven instructions. A classify-agent labels every text with one of seven resilience themes first, so that knowledge extraction and question generation stay anchored to a single topic and avoid cross-domain semantic drift. All generated content is constrained to derive from the input evidence or knowledge units, with prompts and format rules suppressing external facts and essay-style padding.

Training then proceeds in two stages on top of six open-source model families of different sizes and alignment states, including Qwen2.5 models of 7 and 14 billion parameters, Qwen3-8B, Yi-9B, DeepSeek-7B and Mistral-7B. Both stages use low-rank adaptation, or LoRA, in which the frozen original weights are supplemented by a low-rank product of two small matrices, so only a fraction of parameters are trained. Continued pre-training uses a higher learning rate to drive domain knowledge absorption, while supervised fine-tuning uses a lower learning rate to align response formats without disturbing the acquired representations. Training-loss curves showed stable convergence across all six model families, with early-to-final loss reductions of roughly 6 to 26 percent in the pre-training stage and up to nearly 36 percent in fine-tuning, indicating that the generated datasets yield reliable optimization signals rather than divergence.

Evaluation rests on a new benchmark called ResCity-Eval, comprising four progressive tasks: basic knowledge question answering with 1,000 items, mechanism explanation with 700, structured analysis with 700, and strategy and pathway generation with 500. The team introduced a novel metric termed Domain Knowledge Density, which measures the proportion of effective professional information per unit of text by combining domain term hits and information units against length, penalizing the vague, verbose answers that general models tend to produce. Open-ended tasks are scored by an LLM-as-a-Judge setup in which two independent evaluation models, with fixed prompts applied identically to every candidate, identify causal elements, structural organization and strategic components, and their averaged scores feed weighted aggregation.

The headline result came from the final fine-tuned models. Qwen2.5-14B-SFT achieved the highest overall weighted score of 0.683, beating the general-purpose baseline Qwen-max by 0.075 and DeepSeek-v4-flash by 0.099. It led on mechanism explanation, structured analysis and strategy generation, and the strategy gap was especially wide, exceeding Qwen-max by 0.100 and DeepSeek-v4-flash by 0.190 on pathway and governance tasks. An ablation comparison showed the two-stage recipe delivered consistent gains for Qwen2.5-14B, which rose from 0.635 as a base model to 0.683 after fine-tuning, and for DeepSeek-7B, which climbed from 0.275 to 0.494. But the authors are candid that the strategy is not universally beneficial: Qwen3-8B collapsed from 0.648 to 0.308 after fine-tuning, and several other families plateaued or declined, demonstrating that stable training loss does not guarantee downstream improvement and that compatibility between base model and adaptation data matters.

Beyond benchmarks, the researchers built a local intelligent question-answering system and an agent-oriented case study in which Qwen2.5-14B-SFT serves as the language coordinator for seismic resilience assessment of electrical substations. In demonstrations covering lifeline-system resilience assessment, post-earthquake substation task parsing, seismic resilience enhancement strategies and infrastructure indicator design, the model organized professional responses around functional degradation, interdependencies, recovery processes and governance coordination, and converted natural-language requests into structured task states that route engineering tools for vulnerability analysis, functionality evaluation and recovery optimization. The language model deliberately does not replace numerical simulation; it parses intent, plans tool invocation sequences and interprets results produced by external engineering models.

The authors note several limitations that shape how far these findings can be generalized. The current corpus, while broad, should eventually incorporate multilingual, multimodal, real-time and city-specific data sources; the benchmark covers text-based knowledge and reasoning rather than numerical, spatial or scenario-based evaluation; and the agent validation rests on a single substation case rather than a survey across infrastructure systems and hazards. The work is also confined to the training and evaluation protocols described, and the reported scores apply to the specific benchmark tasks rather than to real-world resilience outcomes. Even so, the study charts a template for turning fragmented interdisciplinary knowledge into compact, deployable language models, and points toward integrating them with digital twins, sensor data and human expert review in future resilience planning and disaster-prevention practice.

Subject of Research: Development and benchmarking of domain-specific large language models for urban resilience

Article Title: ResCity-LLM: Building and benchmarking domain-specific large language models for resilient city via a dual-stage data generation pipeline

Article References: Cui, J., Wang, Y., Zhai, C., Li, Y., & Zhang, J. (2026). ResCity-LLM: Building and benchmarking domain-specific large language models for resilient city via a dual-stage data generation pipeline. Results in Engineering, 32, Article 113568. https://doi.org/10.1016/j.rineng.2026.113568

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113568

Keywords: large language models, urban resilience, LoRA fine-tuning, disaster prevention, benchmark evaluation, continued pre-training, supervised fine-tuning, multi-agent data generation, seismic resilience, open-source models, knowledge corpus, intelligent agents