Training-free retrieval-augmented generation with reinforced reasoning for flood damage nowcasting

1. Problem Statement

Flood damage assessment is critical for disaster response but traditionally relies on supervised machine learning models that require extensive task-specific preparation, including dataset assembly, feature engineering, and model training. These workflows are costly, slow to adapt to new regions, and often struggle with highly imbalanced data (where severe damage is rare). Furthermore, standard models often act as "black boxes," lacking the structured reasoning required for explainable decision-making.

The authors propose a solution to these challenges: a training-free framework that leverages Large Language Models (LLMs) to predict Property Damage Extent (PDE) using retrieval-augmented generation (RAG) and reinforced reasoning, avoiding the need for gradient-based fine-tuning during deployment.

2. Methodology: R2RAG-Flood

The proposed framework, R2RAG-Flood, operates on a "reasoning-centric" knowledge base derived from labeled tabular records. It does not train the LLM; instead, it structures the input data and retrieval context to guide the model's reasoning.

A. Reasoning-Centric Knowledge Base

The system organizes historical flood data into a three-level knowledge base:

1. Global References: General patterns across the entire dataset.
2. HUC12-level Prototypes: Examples grouped by 12-digit Hydrologic Unit Codes (watersheds) to ensure hydrologic comparability.
3. Hard-Boundary Cases: Specific records near the decision boundaries (e.g., between "no damage" and "damage," or "medium" and "high" damage) to help the model distinguish subtle ordinal differences.

Each record in the knowledge base is converted into a text-mode summary (a concise, 120-word paragraph) and paired with a reasoning trajectory (a Chain-of-Thought explanation generated by an LLM offline).

B. Feature Engineering & Divergence Analysis

To handle the structured nature of tabular data, the authors perform a divergence-based feature analysis (using Jensen–Shannon divergence and Kolmogorov–Smirnov statistics) to rank predictors based on their ability to separate PDE classes.

• Occurrence Boundary: Separates "No Damage" from "Damage."
• Severity Boundary: Separates "Medium" from "High" damage.
• The top 5 features for each boundary are prioritized in the text summaries and used to select the most relevant "free-shots" (exemplars) for the prompt.

C. Inference Pipeline

The inference process follows a two-stage logic:

1. Context-Augmented Prompting: For a target flood record, the system retrieves:
   • Local Neighbors: Up to 3 labeled records within a 1km radius (providing geographic/hydrologic context).
   • Free-Shots: If local neighbors are sparse, the system injects pre-selected prototypes and hard-boundary cases from the knowledge base.
2. Reinforced Reasoning: The LLM generates a reasoning trajectory (<thought>) followed by a prediction (<answer>). The reasoning is constrained to a two-stage structure: first determining if damage occurred, then determining severity.
3. Conservative Downgrade Check: A rule-based post-processing step analyzes the generated reasoning. If the narrative evidence weakly supports a high-severity prediction (e.g., predicting "High" damage but the text only mentions light cues), the label is downgraded by one level. This reduces false positives for severe damage without introducing an "abstain" class.

3. Key Contributions

1. R2RAG-Flood Framework: Introduction of a training-free, retrieval-augmented generation pipeline specifically designed for ordinal flood damage classification, utilizing a three-level reasoning-centric knowledge base.
2. Flood-Specific Reasoning Pipeline: Development of a hybrid retrieval strategy combining geographically local neighbors with divergence-guided prototypes and hard-boundary cases to support reliable ordinal decisions.
3. Explainable AI: The framework generates structured rationales for every prediction, making the decision process transparent and auditable.
4. Cost-Efficiency Analysis: Demonstration that lighter LLM variants can achieve competitive "severity-per-cost" efficiency compared to heavy supervised baselines and larger LLMs.

4. Experimental Results

The framework was evaluated on a case study of Hurricane Harvey in Harris County, Texas, using 15,378 grid cells (500m resolution) with 14 predictors (built environment, topography, hydrology).

• Baseline Comparison: A supervised gradient-boosted tree model (FloodDamageCast*) achieved the highest overall accuracy (0.714) and damage-class accuracy (0.859).
• R2RAG-Flood Performance: Across seven LLM backbones (including GPT-4o, GPT-4.1, Llama-3.1, Qwen, and DeepSeek-R1), R2RAG-Flood achieved:
  • Overall Accuracy: 0.613 – 0.668.
  • Damage-Class Accuracy: 0.757 – 0.896 (some variants outperformed the supervised baseline in detecting damaged classes).
  • Reasoning Quality: High Label-Reasoning Alignment (LRA) for top-tier models (e.g., GPT-4.1 reached 1.0), indicating the generated explanations matched the predicted labels.
• Ablation Study:
  • Adding local neighbor reasoning provided the most significant performance boost.
  • Divergence-guided free-shots (using the feature ranking) outperformed random-ranking controls, proving the value of the feature selection strategy.
  • The downgrade mechanism improved metrics for stable models (like GPT-4.1) by correcting over-severe predictions but occasionally hurt performance for less stable models.
• Cost Efficiency: Under a "severity-per-cost" metric, lighter models (e.g., GPT-4o-mini, GPT-5-mini) were more cost-efficient than the supervised baseline and larger LLMs, despite slightly lower raw accuracy.

5. Significance and Limitations

Significance:

• Deployment Agility: The framework offers a viable alternative to retraining models for new flood events or regions, as it relies on retrieval rather than parameter updates.
• Explainability: It provides structured, human-readable rationales for damage assessments, which is crucial for emergency management and insurance claims.
• Resource Optimization: It demonstrates that smaller, cheaper LLMs can be highly effective for specific domain tasks when augmented with the right retrieval and reasoning strategies.

Limitations:

• Geographic Generalization: The study is limited to a single event (Hurricane Harvey) in one region; broader generalization requires further testing.
• Reasoning Quality: While automated metrics show alignment, the generated reasoning is not a substitute for expert domain review.
• Retrieval Strategy: The current reliance on geographic proximity (1km radius) may miss feature-similar cases in hydrologically distinct areas; hybrid retrieval strategies are a future direction.

In conclusion, R2RAG-Flood demonstrates that a training-free, reasoning-centric approach can effectively tackle complex, ordinal flood damage prediction tasks, offering a balance of accuracy, explainability, and cost-efficiency that complements traditional supervised learning methods.