DisasterMobLLM: Disaster LLM Frameworks

• DisasterMobLLM is a term for LLM-centered systems in disaster informatics that support both intention-based human mobility prediction and humanitarian information classification.
• It employs techniques like RAG-enhanced intention refinement and LoRA/QLoRA fine-tuning to boost prediction accuracy and cost-effectiveness in disaster scenarios.
• Key studies report significant gains in metrics such as Acc@1, MRR, and F1, demonstrating its robust performance in real-world disaster conditions.

Searching arXiv for DisasterMobLLM and closely related disaster-LLM work to ground the article in current literature. DisasterMobLLM is a name applied in recent disaster-informatics literature to LLM-centered systems for disaster reasoning, but it does not denote a single canonical model. In the narrow sense, it names a disaster-aware human mobility prediction framework that decomposes next-location prediction into intention inference, LLM-based intention refinement, and intention-conditioned location prediction (Tang et al., 26 Jul 2025). In a second, separate usage, it names a lightweight framework for disaster humanitarian information classification from social-media posts using Llama 3.1 8B-Instruct with prompting, LoRA, QLoRA, and optional retrieval augmentation (Jinzhen et al., 21 Jan 2026). In a broader interpretive sense, the surrounding literature uses “DisasterMobLLM-style” to refer to disaster-focused LLM or MLLM pipelines for planning, geocoding, monitoring, simulation, and emergency operations.

1. Terminological scope and disambiguation

The term has at least two concrete referents in the literature, and this ambiguity is central to understanding the topic.

Usage	Core task	Characteristic components
DisasterMobLLM (Tang et al., 26 Jul 2025)	Disaster-scenario next-location prediction	RAG-Enhanced Intention Predictor, LLM-based Intention Refiner, Intention-Modulated Location Predictor
DisasterMobLLM (Jinzhen et al., 21 Jan 2026)	Humanitarian information categorization and event type identification	Llama 3.1 8B-Instruct, constrained JSON prompting, LoRA/QLoRA, optional RAG

The first usage is mobility-centric. It addresses the failure of standard next-location predictors when disasters alter spatial distributions, increase immobility, and invalidate ordinary trajectory regularities. Its basic claim is that disaster mobility should be modeled via mobility intention rather than by directly transferring raw trajectories across cities (Tang et al., 26 Jul 2025).

The second usage is social-sensing-centric. It addresses the extraction of structured crisis intelligence from noisy, imbalanced disaster tweets under resource constraints. Its basic claim is that prompting alone is insufficient, LoRA and QLoRA are the best cost-performance choice, and retrieval-augmented generation is not universally helpful (Jinzhen et al., 21 Jan 2026).

A common misconception is that DisasterMobLLM refers to one standardized disaster foundation model. The literature instead supports a narrower, plural understanding: the name has been reused for distinct systems, while adjacent work extends similar design patterns to geocoding, planning, multimodal monitoring, and agentic emergency response.

2. DisasterMobLLM as disaster-aware human mobility prediction

In "Predicting Human Mobility in Disasters via LLM-Enhanced Cross-City Learning" (Tang et al., 26 Jul 2025), DisasterMobLLM is formulated as a wrapper around existing deep mobility predictors for disaster scenarios. The problem is to predict the next location $r_T$ given a disaster trajectory $S^D_{0:T-1}$, historical normal trajectories $\{S\}^N$, disaster severity $d$, and reference trajectories $\{S\}^{ref}$. A trajectory is represented as $S_{0:t} = \{r_0, r_1, \dots, r_t\}$.

The framework contains three modules. The RAG-Enhanced Intention Predictor converts travel history into intention-oriented features and predicts the next intention embedding. The LLM-based Intention Refiner revises that intention using retrieved reference trajectories and disaster severity. The Intention-Modulated Location Predictor maps the refined intention to an exact next location through a base model such as DeepMove, Flashback, or STiSAN+ (Tang et al., 26 Jul 2025).

Its core methodological move is abstraction from raw trajectory transfer to intention transfer. Travel features are extracted from attributes such as POI composition, distance to transportation facilities, and road-network relations. These are aligned across cities using Transfer Component Analysis and unsupervised clustering, with a separate immobility embedding for unchanged locations. Intention-language alignment is then established through Intention-CLIP, which constructs intention-related prototypes $P = h(V)$ from the LLM vocabulary table and produces a language-space intention embedding

$$T_t = \mathrm{softmax}\left(\frac{X_t P^T}{\sqrt{d_k}}\right) P.$$

The alignment loss is

$$\mathcal{L}_C = \mathrm{InfoNCE}(T_t, Y_t) + \mathrm{CE}(T_t, Y_t).$$

Retrieval is performed over two sources: disaster trajectories from source cities with the same disaster level $d$, and target-city trajectories in normal or other disaster scenarios. Similarity is measured on intention sequences using DTW. The LLM stage uses an intention-incorporated prompt that asks three explicit questions: whether the predicted next intention is correct, whether the next intention should be “stay still,” and, if not, which intention index is correct. Disaster intensity is injected through a soft prompt, with $S^D_{0:T-1}$0, concatenated with the intention prompt embedding $S^D_{0:T-1}$1 as $S^D_{0:T-1}$2 (Tang et al., 26 Jul 2025).

The downstream predictor modifies ordinary next-location prediction by conditioning on intention: $S^D_{0:T-1}$3 where $S^D_{0:T-1}$4 is the base mobility representation and $S^D_{0:T-1}$5 is the intention embedding. The paper states that $S^D_{0:T-1}$6 can be instantiated by element-wise product, concatenation, or attention.

The empirical setting uses 7 real-world trajectory datasets from 7 cities: Qingyuan, Shaoguan, Zhuhai, Wuzhou, Guilin, Hezhou, and Zhongshan, with Zhongshan as the target city. Disaster severity is based on rainfall from CHIRPS, with five categorized rainfall disaster levels. The comparison set includes LSTM, GRU, DeepMove, Flashback, STiSAN+, HMM+MDP, DeepMob, CHAML, CATUS, LLM4POI, and ST-MoE-BERT (Tang et al., 26 Jul 2025).

The strongest reported result is that DisasterMobLLM with STiSAN+ as the base model improves over the best baseline by 32.8% Acc@1, 28.3% MRR, and 35.0% F1@Immob. The best disaster-scenario table entries include Acc@1 = 0.2897, Acc@10 = 0.5118, MRR = 0.3209, and F1@Immob = 0.5062. The motivating degradation is also large: normal-scenario models can lose around 46.4% accuracy and 24.5% MRR when moved to disasters (Tang et al., 26 Jul 2025).

The ablations are equally important for interpretation. Removing retrieval, soft-prompting, explicit immobility handling, or LLM refinement all reduces performance, with the largest degradation on immobility metrics arising when immobility modeling is removed. This suggests that the framework’s gains are not from generic LLM insertion alone, but from the interaction of cross-city retrieval, disaster-level conditioning, explicit “stay still” modeling, and intention-conditioned prediction.

3. DisasterMobLLM as lightweight humanitarian information classification

In "A Lightweight LLM Framework for Disaster Humanitarian Information Classification" (Jinzhen et al., 21 Jan 2026), DisasterMobLLM is a different system: a resource-conscious disaster tweet classifier built around Llama 3.1 8B-Instruct. It defines a dual-task mapping

$S^D_{0:T-1}$7

where $S^D_{0:T-1}$8 is the humanitarian label and $S^D_{0:T-1}$9 is the event type.

The data source is HumAID, normalized into a unified JSONL corpus with 76,484 total tweets, split into 53,531 train, 7,793 dev, and 15,160 test. The humanitarian task has 10 classes and the event-type task has 4 classes: earthquake, fire, flood, and hurricane. The paper emphasizes severe imbalance, including rescue_volunteering_or_donation_effort: 21,278 samples (27.8%) and missing_or_found_people: 358 (0.5%), yielding roughly a 59:1 ratio between the largest and smallest classes (Jinzhen et al., 21 Jan 2026).

The experimental comparison spans zero-shot prompting, few-shot prompting, LoRA, QLoRA, and three RAG variants. The constrained output format is fixed JSON: $\{S\}^N$1 This constraint is necessary because unconstrained prompting often produced schema violations, invented labels, or explanations. All experiments were run on a single NVIDIA GeForce RTX 3090 GPU with 24 GB VRAM, with deterministic decoding using temperature = 0, top_p = 1.0, and max_tokens = 50 (Jinzhen et al., 21 Jan 2026).

Prompting yields modest gains. For humanitarian classification, zero-shot reaches 0.4183 accuracy, manual few-shot 0.3272, static few-shot 0.5051, and dynamic few-shot 0.6410 accuracy, 0.5625 macro-F1, and 0.6534 weighted-F1. For event-type classification, the same methods yield 0.6274, 0.4665, 0.8197, and 0.8569 accuracy, respectively (Jinzhen et al., 21 Jan 2026).

The decisive results come from parameter-efficient tuning. LoRA rank 32 reaches 0.7951 humanitarian accuracy and 0.9875 event accuracy; LoRA rank 64 reaches 0.7962 and 0.9879; QLoRA rank 32 reaches 0.7942 and 0.9875. The paper highlights that rank 64 is only marginally better than rank 32, and that with rank 32 only about 167M / 8B ≈ 2.09% of parameters are trainable. It further reports that QLoRA retains 99.4% of LoRA performance while cutting memory use by 50% (Jinzhen et al., 21 Jan 2026).

A distinctive finding is that retrieval augmentation does not behave uniformly. For the fine-tuned classifier, all tested RAG variants reduce accuracy relative to no RAG. The paper attributes this to semantically similar but label-inconsistent retrieved examples and describes the pattern as a capability-interference trade-off: RAG helps weak models but injects noise into strong ones (Jinzhen et al., 21 Jan 2026).

The paper also argues that some residual error is actually taxonomy ambiguity. On 300 samples from the top confusion pairs, GPT-4 accuracy is 21.67%, and on other_relevant_information vs not_humanitarian it is only 3.33%. The estimated benefit of full GPT-4 correction is limited, improving overall accuracy only from 79.51% to 82.10%, at a cost of about $\{S\}^N$022 for all 1,917 confusion samples (Jinzhen et al., 21 Jan 2026).

4. Relation to adjacent disaster-LLM paradigms

The broader literature shows that DisasterMobLLM is part of a larger shift from narrow disaster classifiers toward operationally structured LLM systems.

One adjacent line is plan generation. "DisasterResponseGPT" generates three candidate plans of action from a scenario description, main objectives, available assets, and main problems, then supports iterative natural-language refinement. In the reported comparison, all three models produced usable plans of action, GPT-3.5 and GPT-4 followed the requested format better than Bard, and DisasterResponseGPT-3.5 was closest to the human plan overall (Goecks et al., 2023).

A second line is crisis communication. "LLM-Assisted Crisis Management" describes two workflows built around LLAMA2: a dispatcher copilot for 911 communication and a public-facing disaster app for large-scale incidents. In its model comparison, LLAMA2-70B-chat achieves Precision 0.82, Recall 0.88, F1 0.85, ROC-AUC 0.93, and Accuracy 0.69, while LLAMA2-13B-chat is selected as the preferred deployment candidate because it balances performance and computational efficiency. The reported inference latency can be brought down to around two seconds (Otal et al., 2024).

A third line is multimodal social sensing. The 3M pipeline for earthquake damage evaluation uses multimodal LLMs to map tweet text and images to Modified Mercalli Intensity. Its headline result is that model-estimated MMI correlates positively with DYFI, reaching r = 0.78 for Qwen on Ridgecrest, while text + image fusion performs best across both studied earthquakes (Ma et al., 3 Jun 2025). "CrisisSense-LLM" similarly reworks disaster tweet analysis as instruction-tuned multi-label classification and reports a best overall accuracy of 0.638 with regeneration and checkpoint ensembling, together with prompt-format sensitivity that can exceed 50% when inference prompts differ from training prompts (Yin et al., 2024).

A fourth line is geospatial grounding and monitoring. "Subnational Geocoding of Global Disasters Using LLMs" presents a fully automated pipeline that geocodes 14,215 EM-DAT disaster events across 17,948 unique locations, corresponding to 92% of entries with location information and 88% of all recorded disasters in the period studied (Ronco et al., 13 Nov 2025). "MONITRS" supplies a multimodal dataset of 9,996 disaster incidents with temporal satellite imagery, news-derived language, geotagged locations, and QA supervision; fine-tuning TEOChat 7B on MONITRS-QA yields 88.69% event classification accuracy and 70.72% temporal grounding accuracy, while location grounding remains much lower at 23.25% (Revankar et al., 22 Jul 2025).

A fifth line is simulation. "LLMs as World Models" treats the LLM as a multimodal reasoning function for pre-event impact simulation and reports Correlation = 0.88 and RMSE = 0.77 at the zip code level against DYFI for earthquake MMI estimation, with RAG, ICL, and visual inputs improving performance (Li et al., 2 Jun 2025). This suggests that DisasterMobLLM-style systems are not confined to post-event triage; they also extend to pre-event synthetic sensing.

5. Methodological tensions and recurrent misconceptions

The literature reveals several tensions that recur across systems using the DisasterMobLLM label or analogous architectures.

The first concerns retrieval. In disaster mobility prediction, retrieval is structurally beneficial: removing RAG reduces performance because the model loses external cross-city examples and therefore weakens intention refinement under disaster conditions (Tang et al., 26 Jul 2025). In lightweight humanitarian classification, by contrast, RAG degrades a strong fine-tuned model because retrieved neighbors are often semantically similar but label-inconsistent (Jinzhen et al., 21 Jan 2026). The methodological lesson is not that retrieval is either good or bad in general, but that its value depends on whether the task needs external cross-context analogues or strict label-consistent classification.

The second concerns modality. A common assumption is that disaster LLM work is primarily textual. The adjacent literature contradicts this. MONITRS is built around temporal satellite imagery and natural-language annotations; the 3M pipeline depends on image-text fusion; the pre-event world-model study reports that street-level imagery is the clearest contributor to improvement; and geocoding pipelines use LLMs primarily as text-to-structure engines rather than as classifiers (Revankar et al., 22 Jul 2025, Ma et al., 3 Jun 2025, Li et al., 2 Jun 2025, Ronco et al., 13 Nov 2025). This suggests that the stronger family resemblance is not “tweet classification,” but heterogeneous evidence integration.

The third concerns operational readiness. "Can LLM Agents Respond to Disasters?" introduces DORA, with 515 expert-authored tasks, 45 real-world disaster events, 3,500 tool-call steps, and a 108-tool MCP library. Even the best model, Gemini-3.0-Flash, reaches only 53.74% average accuracy, while the gold trajectory reaches 80.48%. The paper identifies three persistent disaster-specific failure modes: damage-semantic grounding — 20.3%, sensor-modality mismatch — 14.8%, and disaster-pipeline composition — 56.3%. It also reports that the agent-to-gold gap widens from about 7% on short pipelines to 56% on trajectories with 11 or more steps (Wang et al., 12 May 2026). The misconception that disaster-response LLMs are already end-to-end operationally reliable is therefore not supported.

6. Limitations, unresolved issues, and likely research trajectory

The two systems named DisasterMobLLM inherit different bottlenecks. The mobility-prediction framework depends on intention abstraction, cross-city transfer, and disaster-level prompting; its gains are substantial, but they are evaluated in a rainfall-conditioned setting with one explicit target city, and the framework is described as slower than some smaller LLM-based baselines, though still in the same magnitude for deployment-oriented settings (Tang et al., 26 Jul 2025). The tweet-classification framework is lightweight and reproducible, but its residual errors concentrate in ambiguous classes such as other_relevant_information, not_humanitarian, and requests_or_urgent_needs, indicating that taxonomy design remains a core bottleneck rather than a purely modeling bottleneck (Jinzhen et al., 21 Jan 2026).

The adjacent literature points to additional system-level constraints. DisasterResponseGPT has no image-based input support, operates under context-window limits of about 4096 or 8192 tokens, and cannot generate functional visual plan sketches (Goecks et al., 2023). MONITRS is U.S.-only, relies on RGB Sentinel-2 imagery, and remains weak on location grounding (Revankar et al., 22 Jul 2025). The pre-event world-model study evaluates only two earthquake cases, both in California (Li et al., 2 Jun 2025). DORA shows that even frontier agents remain fragile in long, heterogeneous geospatial workflows (Wang et al., 12 May 2026).

A plausible implication is that the research trajectory is moving toward increasingly structured, multimodal, and operation-aware disaster LLM systems. One branch emphasizes intention-level transfer and immobility modeling for mobility forecasting (Tang et al., 26 Jul 2025). Another emphasizes parameter-efficient adaptation under strict resource budgets (Jinzhen et al., 21 Jan 2026). A third emphasizes tool-grounded geospatial reasoning, temporal monitoring, and multimodal disaster simulation (Wang et al., 12 May 2026, Revankar et al., 22 Jul 2025, Li et al., 2 Jun 2025). Taken together, these directions suggest that DisasterMobLLM is best understood not as a single model family with a fixed architecture, but as a converging research theme: the use of LLMs and MLLMs to convert heterogeneous disaster signals into structured operational inferences under severe time, data, and deployment constraints.