🤖 AI

What Will Happen Next: Large Models-Driven Deduction for Emergency Instances

This paper proposes the Large Models-driven World Line Divergence System (WLDS), a novel framework that leverages large models with factual and logical calibration mechanisms to dynamically generate, visualize, and deduce diverse, high-fidelity emergency instances for improved risk assessment and decision-making.

Original authors: Zhengqing Hu, Dong Chen, Junkun Yuan, Liang Liu, Hua Wang, Zhao Jin, Yingchaojie Feng, Wei Chen, Mingliang Xu

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhengqing Hu, Dong Chen, Junkun Yuan, Liang Liu, Hua Wang, Zhao Jin, Yingchaojie Feng, Wei Chen, Mingliang Xu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict what happens next in a chaotic situation, like a fire breaking out on a subway platform or a car losing control on a highway. Traditionally, computers try to do this by following a strict rulebook, like a train on a single track. If the rulebook doesn't have a specific rule for "fire in a trash can," the computer gets stuck or makes up something that looks real but is actually impossible (like a fire truck driving onto a subway platform).

This paper introduces a new system called WLDS (World Line Divergence System). Think of WLDS as a super-smart storyteller who doesn't just follow a single script but can imagine many different "what-if" stories, while still sticking to the laws of physics and logic.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Hallucination" Trap

The authors point out that if you just ask a powerful AI (a "Large Model") to guess what happens next in an emergency, it often gets two things wrong:

Factual Deviation (The "Magic" Problem): The AI might say a car with 5% battery can drive 500 miles. It's a cool story, but it breaks the laws of physics.
Logical Deviation (The "Plot Hole" Problem): The AI might say a car is on a busy highway, and then suddenly, in the next sentence, it encounters a pedestrian crossing that doesn't exist on highways. The story makes no sense.

2. The Solution: The "World Line" Concept

The authors use a concept from physics called a "World Line," which is like a path an event takes through time and space.

The Setup: Instead of guessing one future, WLDS imagines multiple possible futures (like a "Choose Your Own Adventure" book).
The Magic Trick (Cross-Domain Transfer): Since real emergency data is rare (fires are bad, so they don't happen often), the system learns from other places. It takes lessons from, say, a chemical plant fire and uses that knowledge to imagine what a subway fire might look like. It's like a chef using a recipe for a steak to figure out how to cook a fish.

3. The "Double-Check" Mechanism

This is the most important part. Before the AI tells you the story, it runs it through two strict filters:

The Fact-Checker: It looks at a library of real-world facts (like "fire trucks can't fit in subway tunnels"). If the story says a fire truck enters a tunnel, the Fact-Checker says, "Nope, that's impossible," and rewrites that part.
The Logic-Checker: It looks at the flow of the story. If Step 1 says "the fire is out" and Step 2 says "people are evacuating because of the fire," the Logic-Checker says, "Wait, if the fire is out, why are they running?" It fixes the timeline so the story makes sense.

4. The Visual Storyboard

Once the story is checked and approved, the system doesn't just give you text. It creates a visual storyboard. It matches the text to real images or generates new pictures that perfectly match the scene. So, if the text says "smoke is filling the station," you see a picture of a smoky station, not a sunny park.

What Did They Prove?

The team built a test dataset called EID with 4,300 different emergency scenarios across 10 fields (like autonomous driving, subway systems, and nuclear plants).

The Results: When they tested WLDS against other AI models, WLDS was much better at telling stories that were both factually true (no magic batteries) and logically sound (no plot holes).
Expert Approval: They showed the results to 20 human experts (like safety engineers). The experts gave the system very high scores, especially in complex fields like autonomous driving, saying the stories were realistic, diverse, and useful for planning.

In a Nutshell

WLDS is a tool that helps us simulate emergencies safely. It acts like a creative writer who is also a strict editor. It can imagine many different ways a disaster could unfold to help us prepare, but it constantly checks its own work to make sure nothing impossible or illogical slips through the cracks. This helps safety experts train for the worst-case scenarios without ever having to experience the real danger.

Technical Summary: What Will Happen Next: Large Models-Driven Deduction for Emergency Instances

1. Problem Statement

Traditional simulation technologies for emergency instances face two critical limitations:

Lack of Randomness and Diversity: Existing systems rely on preset rules and static mappings, failing to model the stochastic nature of event states and the diverse development paths found in the physical world. This is particularly problematic for high-risk, low-frequency domains (e.g., urban rail transit, autonomous driving) where emergency data is scarce.
Hallucinations in Large Models (LMs): While LMs offer dynamic generation and cross-domain knowledge transfer, directly applying them to emergency deduction leads to two specific types of hallucinations:
- Factual Deviation: Generated content violates physical laws or domain specifications (e.g., a vehicle with 5% battery claiming 500km range).
- Logical Deviation: The deduction process suffers from causal disconnections and broken element consistency (e.g., a vehicle on a highway encountering a pedestrian crosswalk).

Current methods struggle to balance the need for diverse, randomized scenarios with the strict requirements of factual accuracy and logical rigor necessary for emergency decision-making.

2. Methodology: The World Line Divergence System (WLDS)

The authors propose WLDS, a simulation and deduction system driven by Large Models that integrates cross-domain knowledge transfer, controlled randomness, and a dual calibration mechanism. The system operates through four core steps:

A. Knowledge Transformation

To address data scarcity, WLDS leverages the cross-domain transfer capabilities of LMs. It constructs an accident dataset ( $D_{acc}$ ) from various domains and uses tailored prompts to generate initial emergency instances for a specific target domain ( $B$ ). This allows the system to migrate emergency knowledge from data-rich domains to data-scarce ones.

B. World Line Deduction

Starting from an initial instance, WLDS generates multiple potential future scenarios (world lines). It introduces controlled randomness via a temperature parameter ( $\tau_k$ ) to balance logical plausibility with deduction diversity. Users can interactively select a specific deduction direction, creating a multi-branch structure where different development paths are explored.

C. World Line Calibration (Dual Mechanism)

To mitigate hallucinations, WLDS employs a dual calibration mechanism after initial generation:

Factual Calibration: Ensures individual events align with a domain-specific knowledge base ( $K_B$ ). The system retrieves relevant facts and computes a consistency score. If the score falls below a threshold, the event is revised to ensure physical and domain accuracy.
Logical Calibration: Evaluates the causal and temporal relationships between consecutive events. A logical discriminator checks for consistency; if a logical flaw is detected (e.g., causal disconnection), the subsequent event is regenerated to restore logical rigor.

D. World Line Visualization

To enhance interpretability, WLDS integrates text and images. It matches generated text descriptions against a keyframe image library. If no suitable image exists, a text-to-image model generates a new one. A text-image alignment discriminator filters out images with significant semantic deviations, ensuring the visual output accurately reflects the deduced narrative.

3. Key Contributions

Systematic Analysis of LM Limitations: The paper identifies and categorizes the specific "factual" and "logical" deviation issues when using LMs for emergency simulation.
WLDS Framework: The proposal of a system that combines LMs with factual and logical calibration mechanisms to achieve high-precision, high-fidelity emergency deduction.
EID Benchmark Dataset: The construction of the Emergency Instances Deduction (EID) benchmark, comprising 10 sub-datasets (e.g., carrier-based aircraft, urban rail transit, chemical plants) with 4,300 three-step branched samples annotated with expert labels, probability, and loss severity.
Automated Evaluation Mechanism: The design of quantitative metrics for Factual Consistency (FC) and Logical Consistency (LC) to objectively evaluate simulation performance.

4. Experimental Results

Experiments were conducted on the EID benchmark across 10 domains, comparing WLDS (integrated with various LMs like Hunyuan, GLM, and Qwen) against baseline models.

Performance Gains: WLDS demonstrated superior performance in both FC and LC. In the Urban Rail Transit domain, WLDS combined with Hunyuan-Turbos achieved an FC of 93.75% and LC of 91.67%, representing improvements of 7.08% and 8.34% over the baseline, respectively.
Scenario Prediction Accuracy: In the Chemical Plant sub-dataset, WLDS achieved a scenario prediction accuracy of 91.50%, an 8.50% improvement over the baseline.
Ablation Studies: Removing either factual or logical calibration resulted in significant performance drops, confirming that both mechanisms are essential for mitigating hallucinations and ensuring high-fidelity deduction.
Expert Evaluation: Domain experts rated WLDS highly across diversity, precision, rigor, and feasibility. The Autonomous Driving domain received the highest average score (4.8/5), indicating strong recognition of the system's ability to generate diverse, accurate, and logically rigorous scenarios.

5. Significance and Claims

The paper claims that WLDS provides a novel approach to emergency simulation by overcoming the rigidity of traditional rule-based systems and the unreliability of raw LMs. Its significance lies in:

Data Scarcity Mitigation: Enabling the generation of realistic emergency instances in domains where historical data is rare through cross-domain knowledge transfer.
Decision Support: Providing a high-fidelity, interpretable tool for risk assessment and emergency decision-making, capable of visualizing diverse risk evolution paths (from low-risk containment to high-risk stampedes).
Interpretability: Moving beyond text-only generation to a text-image fused visualization that aligns with domain semantics, making complex deduction chains accessible to human operators.

The authors acknowledge limitations, noting that logical coherence may decrease in very long-chain deductions and that current visualization lacks dynamic temporal rhythms. Future work aims to incorporate multi-stage deduction mechanisms and dynamic video generation to address these gaps.